characterizing copper nanoparticle dc sputtering with...
TRANSCRIPT
Characterizing copper nanoparticle DC sputtering
with oxygen interference
Arjan Bijlsma (s1398245)
February 14th - December 1th 2011
Abstract
The properties of a material are directly dictated by its size, shape and internal structure. In a time where
research on nanoscale is widely practiced and the demand for the smallest possible electronics grows, new
challenges emerge in the creating and usage of nanomaterials. In this thesis the workings of a new Nanosys550
Deposition System from MANTIS Deposition LTD for creation of nanoparticles is described with in the present
work a specific focus on copper clusters. First the basics of direct current (DC) magnetron sputtering are
explained. This form of sputtering works by accelerating argon plasma against a copper target, thereby
breaking of single copper atoms which form a gas. By cooling this gas into supersaturation, copper nanoclusters
are formed by condensation processes. Although, these processes are controlled by only a few system settings,
together they comprise a complex system. These settings and their effects are discussed and have been
investigated by experimentation. The latter has been done in a first experiment series by Half Factorial Design
(HFD) with Center Point Experiments (CPE) to characterize the standard operation window of the new system.
A second experiment series on Energetic Cluster Impact (ECI) has been performed to examine the charge
character of the sputtered copper clusters by attempting cluster acceleration with an applied bias voltage to
the substrate holder. The copper clusters were deposited on TEM grids and silicon wafer pieces for both
experiment series. The results of the HFD experiments were imaged with the use of a Transmission Electron
Microscope (TEM) and analyzed by statistical software, which calculated the effects of the separate system
variables. Assessment of these effects revealed that they behaved according to a-priori (qualitative)
expectations but were non-significant. This led to the discovery of oxygen leakage into the vacuum system and
several attempts to prevent this problem. At the end, the oxygen problems have been solved and the observed
HFD results are explained by literature research. Simultaneous with the preventive measures the ECI
experiments were performed. The results were imaged also by TEM and an Atomic Force Microscope (AFM).
The obtained data from both imaging techniques revealed no support or proof for ECI, although the absence of
cluster charge seems unlikely from literature research.
Master thesis for the study of Applied Physics
Student: Bsc. Arjan Bijlsma (s1398245)
Supervisor: ing. G.H. ten Brink
Group leader: prof. dr. ir. B.J. Kooi
Dept. Nanostructured Materials and Interfaces
Zernike Institute for Advanced Materials
University of Groningen
Nijenborgh 4
9747 AG Groningen
The Netherlands
2 / 38
Contents
Overview ..................................................................... 2
1. Introduction: opportunities on nanoscale .............. 3
2. Theory: Inert Gas Condensation with sputtering .... 3
2.1. DC magnetron sputtering ................................. 4
2.2. Supersaturation for condensation ................... 5
3. Theory: clusters ....................................................... 5
3.1. Nucleus formation ............................................ 6
3.2. Cluster growth .................................................. 6
4. Theory: gas properties and system variables .......... 7
4.1. Important gas properties ................................. 7
4.1.1. Gas species and charge ............................. 7
4.1.2. Gas ratios................................................... 8
4.1.3. Mean free path ......................................... 8
4.1.4. Dwell time ................................................. 8
4.2. System variables ............................................... 8
4.2.1. Partial pressures ........................................ 9
4.2.2. Magnetron strength and power ................ 9
4.2.3. Temperature ............................................. 9
4.2.4. Aggregation length .................................. 10
4.2.5. Bias voltage ............................................. 10
4.3. Simulations and magic numbers .................... 10
5. Experimental setup ............................................... 11
5.1. Equipment ...................................................... 11
5.2. Settings: Half Factorial Design ........................ 12
5.3. Settings: Energetic Cluster Impact ................. 14
6. Results .................................................................. 14
6.1. Imaging .......................................................... 14
6.1.1. Transmission Electron Microscope ......... 14
6.1.2. Atomic Force Microscope ....................... 16
6.2. Results: Half Factorial Design ........................ 16
6.2.1. Image Quantification Analyses ............... 16
6.2.2. Statistical Data Analyses ......................... 18
6.3. Results: Energetic Cluster Impact .................. 19
6.3.1. TEM images of high/low bias voltages ... 19
6.3.2. TEM and AFM images ............................. 21
7. Discussion ............................................................. 23
7.1. Discussion: Half Factorial Design ................... 23
7.1.1. Factor effects .......................................... 23
7.1.2. Significance of factor effects .................. 24
7.2. Discussion: Energetic Cluster Impact ............ 25
7.2.1 Bias voltage effects .................................. 25
7.2.2 Combining TEM and AFM ........................ 26
7.3. Oxygen interference ...................................... 29
8. Conclusions ........................................................... 31
9. References ............................................................ 32
Appendix A.1. Experiment design ............................ 34
A.1. Factorial Design (FD) ..................................... 34
A.2. Half Factorial Design (HFD) ........................... 36
A.3. Center Point Experiments (CPEs) .................. 37
Acknowledgments .................................................... 38
Overview
This report will start with a general introduction on the field of nanoscience and its potential opportunities in
chapter 1. In chapter 2 the reader will be familiarized with the theory behind the creation of nanoparticles by
inert gas condensation with a DC magnetron sputtering system. Further details about the growth process of
nanoclusters will be explained in chapter 3. Hereafter, chapter 4 gives an overview of the gas properties which
are of key importance for cluster growth and how the available system variables influence them. In chapter 5
the equipment and settings of the first experiment series by Half Factorial Design (HFD) and the second
experiment series on Energetic Cluster Impact (ECI), will be described. Detailed information on the settings for
a Half Factorial Design can be found in appendix A. Chapter 6 will present all results and describes the used
analyses software and imaging techniques. The significance of the results for both experiment series and
discovered oxygen interference is discussed in chapter 7. The final conclusions and recommendations are
summarized in chapter 8.
3 / 38
1. Introduction: opportunities on nanoscale
The field of material science started with the growing interest for non-metallic materials like ceramics,
polymers and semiconductors, in the 1960s [1]. Materials science is generally based on understanding the
relation between structures and properties, where the third and most important element, to tune structures
and properties, is processing. Nowadays, the advances in material science lie primarily in understanding the
structure-property relations for materials with structural control down to the nanometer and atomic scale and
where properties can be strongly size-dependent in the nanometer size range. Important drivers for these
developments have been the increasing capabilities of transmission electron microscopy (TEM) invented by
Ruska et al. in the 1930s, which became widely available in the 1960s, and scanning tunneling microscopy
(STM) and atomic force microscopy (AFM) by Binning et al. in the 1980s. Because of these developments an
even smaller scale, in the nanometer range, opened up for research. This so-called nanoscience concerns the
field of materials and applications in which the typical structures have a size of 1 to 100 nanometers (a billionth
of a meter). This enabled the research of bottom-up produced single entities like clusters [Chap.3.1.],
molecules or even atoms and top-down produced structures often employing lithographic techniques in
combination with etching.
An interesting phenomenon in this size regime is that previously theoretically described quantum mechanical
effects can be observed as important properties of nanometer sized materials. Where the behavior of
macroscopic bulk material are best described by the rules of classical physics and dynamics, nanometer
particles lose some degree of this continuous character and begin to exhibit discrete, more quantum-like
properties. A major part of this change is caused by the energy levels of particularly electronic states, of those
nanoparticles [2]. Also, in descending to the nanoscale the surface of the particles become increasingly
important. For instance, a copper particle of 13 atoms constitutes the smallest stable copper cluster. This
number therefore is the first so-called ‘magic number’. It consists of only one central atom surrounded by 12
neighbors. A one shell bigger copper cluster of 55 atoms, the second ‘magic number’, still has approximately 32
atoms (58 %) on its surface. Even for particles containing 10,000 atoms still nearly 20 % are surface atoms [3,
4]. This will greatly amplify any behavior concerning surface properties and chemical reactivity, like for instance
in the melting point or properties related to catalysis. The above exposes the dissimilarity in the branch of
nanoscience with respect to the overall materials science: at nanoscale, materials can exhibit non-bulklike
behavior, the latter sliding from continuous to discrete when decreasing in size.
This discovery raised a newfound interest in materials like semiconductors and metals. Semiconductor
nanostructures already below 50 nm show quantized atomic-like behavior and are often referred to as
quantum dots [5]. For most metals, particles sizes have to go down to clearly below 10 nm in order to arrive at
discrete atomic-like energy levels. Because of this possible quantized behavior, metals can also potentially work
as nanoscale semiconductors [3], opening up opportunities in different fields of science like computing, optics,
magneto-electronics [6] and medical technology [7].
2. Theory: Inert Gas Condensation with sputtering
A common method for creating nanoparticle is by inert gas condensation (IGC) and has been known since the
1930s [8]. Although there are several ways to accomplish IGC it always involves a method for vaporizing a
target material and creating a supersaturated vapor. If then some of the vapor particles join within the
supersaturated state, referred to as seed or just nucleus, they can acts as a starting point for growth. Under the
right conditions they grow to become nanoparticles, also called nanoclusters. A flow of these nanoclusters can
then be deposited on a desired substrate for analysis. These processes are all preformed in a system which is
pumped to high vacuum prior operation to remove and prevent contamination with other gasses.
Nanocluster growth in a gas phase with IGC has some advantages in contrast to the alternatives, like growth
directly on a (cooled) substrate. The conditions for maintaining a homogeneous gas state for growth are easily
4 / 38
controllable. Also, there are no accumulating effects which can distort the cluster growth because of the
continuous nature of IGC. During operation there is a continuous creation and flow of supersaturated vapor
and therefore cluster growth, in the growth region. This throughput prevents any buildup of detrimental
effects, which can be an unavoidable risk with growth directly on a substrate [1, 8].
As mentioned, there are several ways for vaporizing the desired material for IGC. The use of a resistive filament
is the oldest and still in use method [9]. Currently also techniques with the use of crucibles [10, 11], electron
beams, lasers and magnetron sputtering discharge are used worldwide. The latter was employed during this
research, because of its flexibility in choice of materials, among others. Magnetron sputtering is not restricted
to metals only, as with thermal evaporation for example. Another advantage of magnetron discharge is its high
particle yield because of the uniformity of the discharge current at a large target material surface [1, 12].
2.1. DC magnetron sputtering
A magnetron sputtering system is relatively simple and consists of five main components; an anode (1)
mounted above a target material disk (2) which is placed on top of a magnetron (3) under gas flow (4) in a
cooled volume (5). The first three components together are called the magnetron head [see figure 2.1].
The process of magnetron sputtering starts by supplying one or more gasses into the system from behind the
magnetron head. These gasses will have three functions throughout the system: first sputtering, then
transporting and finally cooling of the target material vapor. With respect to a pure end result noble gases are
used for this purpose, since their inert properties cause minimal interference with the target material. For
carrier gasses the common choice of argon and helium was made in this research. These can be labeled
separately according to their selection for primary functionality. Here, argon was used primarily for sputtering
because of its easy ionization and is therefore called the sputtering gas. The high thermal conductivity of
helium make it more suitable for cooling and thus is labeled cooling gas. Since both play a role in the
transportation of the target material vapor through the system, both are called carrier gasses. It is also possible
to run the system with only the sputtering gas. This will decrease the cooling capabilities, which makes it
harder to accomplish the required supersaturated vapor state later on. Because of this disadvantage the latter
method was not used during this research.
For the sputtering process the supplied sputtering gas has to be ionized above the magnetron head to create a
plasma. This is accomplished by applying a potential between the target material disk, acting as cathode, and
the anode which is on top of the target material. The potential can be alternating current (AC) or direct current
(DC), called RF (Radio Frequency) or DC sputtering, respectively. Although RF sputtering also enables the use of
insulating materials, DC sputtering is a less complex process and therefore better understood. Because of this
Figure 2.1: schematic of sputtering unit.
5 / 38
advantage and the sole interest in metals, DC
sputtering was used during this research. When a
large enough DC potential is applied, electrons are
accelerated away from the target material and
collide with the sputtering gas which passes over
the magnetron head. This causes a discharge which
ionizes the sputtering gas into excited species,
called the primary plasma. The primary plasma can
be seen by the human eye because of its decay to
lower energy levels by emission of visible light. The
frequency of this light is characteristic for the
plasma’s elemental composition [1, 8]. Finally, the
primary plasma is accelerated by the DC potential
and impacts on the target material disk, thereby
breaking off particles which together compose the
desired target material vapor. This process is called
sputtering.
Upon impact of the sputtering ions on the target, secondary electrons are released back into the sputtering gas
cooperating to extra ionization of the primary plasma. To increase the efficiency of the sputtering process a
strong magnetron is placed behind the target material to further enhance the primary ionization. Because of
the magnetic field of the magnetron, the free electrons will travel in a helical path through the plasma. This
increases their effective path length and therefore their chance of colliding with the sputtering gas. The
magnetic field also confines the electrons and plasma above the target which intensifies the sputtering process
even more. The confinement and helical path can easily be seen on the target material disk which, after some
use, shows a torus shaped groove called ‘race track’, revealing the local magnitude and shape of the sputtering
process [see figure 2.2] [1, 4, 12].
2.2. Supersaturation for condensation
The target material vapor, after it’s created by sputtering, is swept to the aggregation volume just above the
primary plasma. Here, the vapor is confined by the local inert gas pressure (of tens or hundreds Pa) and
cooling, to become supersaturated [1, 8, 13]. The relatively high inert gas pressure, which is created by the
small opening of the exit orifice at the end of the aggregation volume, decreases the diffusion rate of the vapor
atoms. The cooling aids this confinement by draining the thermal and kinetic energy from the vapor particles,
thereby further limiting their diffusion rate. The cooling of the vapor is accomplished by cooling the walls of the
aggregation volume. This is enhanced by the heat transfer of the cooling gas by its high thermal conductivity
[14, 15]. Together, the pressure and cooling limit the vapors mean free path which leads to the
supersaturation. In this supersaturated vapor state nuclei can form and start growing to become clusters.
3. Theory: clusters
In the field of Nanotechnology, or more specific Condensed Matter, a particle is called a cluster when it consists
of 2 – 10n atoms, where n can be as high as 6 or 7 [16], and forms the intermediate size between molecules and
bulk material. As mentioned before, clusters can have astonishing intrinsic properties because of their large
surface to size ratio and size-dependent electrical behavior. Also, clusters are the basic building blocks in thin
films, another growing branch in Nanotechnology. A basic process that leads to the creation of clusters is
something we see in everyday life, for example in smoking fires or fog formation. In all cases a substance is
vaporized into a colder volume where the vapor particles start to pack together into clusters [4]. As described
before [Chap.2.], this is a two stage process: first the formation of a nucleus and second the subsequent growth
to become a cluster.
Figure 2.2: copper target material disk, with ‘race track’ groove.
6 / 38
3.1. Nucleus formation
For a cluster to start growing, a nucleus or nucleation center is required. This is a small particle consisting of
only a few monomers. In the case of magnetron sputtering, these monomers are single atoms sputtered from
the target material disk [4, 8, 10, 12, 13]. For these single atoms to join, the high thermal and kinetic energy
from the sputtering process has to be removed. This is accomplished by cooling the vapor through admixing of
the cooling gas as described above. The cooling gas will drain energy from the vapor atoms by inelastic
collisions (collisional cooling) [17]. When cooling is done sufficient, formation of dimers, i.e. two connected
atoms, can start by stable three-body collisions. This is a relatively slow process as three atoms have to collide
simultaneously, thereby forming the bottle neck of the overall process. The collision of two particles would not
lead to a stable dimer [1, 4], as the third atom needs to remove the excess of internal energy for the binding
process. These new dimers will statistically grow and decay by competition of accretion and evaporation of
atoms. Evaporation of atoms will occur because it redeems energy. The early nuclei are continuously heated by
the binding energy of accreting atoms (latent heat of condensation), but cannot distribute this energy
internally because of their small size. Therefore, evaporation can be used as a cooling mechanism.
This competition will continue until a small cluster of critical size r* is formed, given by
the formula 3.1. Where it is assumed that isotropic mesoscopic porperties hold. In the
formula σ stands for the surface energy of a small droplet of atoms, ρ its density, m its
atomic or molecular mass and φk= pk/ps its condition for supersaturation (pk and ps are
the vapor pressure and saturation vapor pressure, respectively). For larger radii
accretion of atoms on the small cluster becomes thermodynamically favorable and
growth becomes a very rapid process. This occurs because the cluster size permits the transfer of the added
condensation energy per accreted atom to its internal degrees of freedom [18]. Although the clusters grow a
little by this condensation, the partial pressure of the vaporized material is rapidly depleted. This quickly
reduces the supersaturation and quenches additional nucleus formation, putting a hold to this brief but rapid
nucleation burst [14].
The previous process of nucleation can best be described theoretically by Classical Nucleation Theory (CNT)
[10], which assumes that nuclei grow and decay by attachment or detachment of one atom at the time
respectively and where for subcritical nuclei detachment is favored over attachment, which is reversed for
supercritical nuclei. CNT is a thermodynamic approach that leads to a barrier in the free energy which reveals
the critical cluster size of n* atoms (instead of radius r*), referring to the smallest stable cluster that allows
subsequent growth with a decrease in free energy [1, 10]. These critical clusters will typically consist of 7-50
atoms depending on temperature and pressure. Below this size the cluster is believed to decompose back into
the vapor. For high supersaturation CNT is expected to fail, because the gas state changes faster than the time
to establish a local metastable equilibrium. CNT also assumes that cluster properties only depend on cluster
size n or r, and therefore there is a single nucleation path. In this way CNT corresponds to a 1D Markov chain.
In practice, clusters can have various shapes and energies for the same cluster size and then there are multiple
nucleation paths [19]. Furthermore, CNT is engineered only for nucleus formation and is not suited for the
subsequent stage of cluster growth. Accordingly, CNT gives no details about the cluster size and shape
distribution in the end result.
3.2. Cluster growth
After the nucleation burst, when all atoms have aggregated into small clusters, growth will continue through
the joining of clusters by Brownian coagulation. Meaning that as clusters collide, they stick together and form
bigger clusters. The name stems from the equal coagulation process known for liquids [14]. Once these
mechanisms are too weak or forbidden, e.g. because of Coulomb repulsion of charged clusters, there is one
more equilibrium that can set. This equilibrium, called coalescence or Ostwald ripening, follows from the
equilibrium between clusters and free atoms and depends on the atomic vapor formed by evaporation and
attachment of atoms from and to the cluster surface. For small clusters evaporation dominates, leading to the
Formula 3.1: critical cluster size [4]
7 / 38
disintegration of these clusters into cluster fragments and atoms. For
big clusters this is reversed; the free fragments and atoms will
coagulate and accrete onto the big clusters. In this way the cluster
density in the final deposition will decrease, but the cluster size will
increase [4].
Initially, the gas temperature is high enough to supply the required
energy to the clusters for remaining spherical with coalescing
fragments, before additional particles collide. As the clusters remain in
the aggregation volume even longer, they will be cooled down further.
Eventually, this will limit the energy for coalescence and thereby the
energy for neck formation between clusters. From that point on
clusters will only combine by agglomeration, meaning that big clusters
stick together with only a small surface area. These combined particles
develop in a dendritic fractal-like structure. For research on the size
and shape distribution of single nanoparticles, this can be an
undesirable effect [14].
For all processes described above one has to keep in mind that they all
happen in the same aggregation volume in rapid succession. Since this
aggregation volume is relatively short, the different gas phases and
accompanying reactions, have a certain overlap with mixed states.
Therefore the above described processes are not restricted by specific
regions and can occur side-by-side. For example an atom evaporated
during coalescence can theoretically contribute to a three-body
collision, although statistically unlikely.
After the growth processes the final clusters exit the aggregation volume with the carrier gasses through a
small exit orifice into the deposition volume. Here a substrate can be mounted for deposition of the clusters.
4. Theory: gas properties and system variables
To create favorable conditions for cluster growth several variables in the system can be varied: gas choice and
flow rates, sputtering power, magnetron strength, temperature, aggregation length and volume geometry.
Although all of these variables together create a complex system, they only influence four important properties
of the gas mixture: species, gas ratios, mean free path and dwell time. Together, these properties dictate the
process and therefore the final result: the species prescribe the possible chemical reactions and composition;
the gas ratios provide the availability of the different species; the mean free path indicates the reaction rates
and the dwell time is a measure for growth time. The relations between the properties and variables have been
described extensively in previous research on nanoparticle creation by magnetron sputtering and reveal the
validity the described theory.
4.1. Important gas properties
To understand what happens during sputtering and how the change of a system variable affects the final result,
the following properties have to be considered separately.
4.1.1. Gas species and charge
The first property of the gas mixture, i.e. species, is obviously decided by the desired clusters in the end result.
But besides the growth reaction described in the previous chapter [Chap.3.], other important processes can
occur in real sputtering systems. What has to be considered is what other species of the present materials can
Figure 3.1: illustration of processes for nanoparticle synthesis [14].
8 / 38
be created during the energetic sputtering process and all possible reactions between them. The creation of
excited or ionized particles from the target material can considerably influence the growth of the clusters, e.g.
by repulsion forces. As sputtering is a highly energetic process, a large fraction of the created clusters is ionized
because of the many charged and electronically excited species present in the afterglow from the sputter
discharge, acting as very efficient ionizers. For sputtered metal atoms only about 10-3
is ionized, still if only one
ionized atom combines with a neutral cluster, the total cluster becomes ionized. Other examples of
mechanisms for ionizing the clusters are by charge transfer from the ionized sputtering gas nearby the cathode
or by impact of an electron from the sputtering process. For copper, it has been found experimentally that 20
to 50 % has a negatively charge [1, 8, 15]. These ionized clusters are suspected of causing some of the large
differences observed between simulation and real experimental results [1]. Furthermore, the choice of
appropriate carrier gasses is of paramount importance for the purity and size of the final clusters. The inert
character of noble gasses and their purity will prevent undesirable reactions, avoiding contamination of the
final clusters [12]. Additionally, it has been observed that heavier carrier gas atoms are more effective in
limiting the mean free path of the target vapor, as obtained particles increased in size when going from helium
to argon to xenon gas [20].
4.1.2. Gas ratios
The second property of importance is the ratio between the supplied carrier gasses and the created target
vapor inside the aggregation volume. Typically, these ratios are referred to by their partial pressures of the
individual gasses, although physically difficult to measure separately. The importance of the partial pressures
results from the fact that every gas in the system has a different role in the sputtering and aggregation
processes. Different ratios can give rise to the same total pressure, but a completely different end result [14].
4.1.3. Mean free path
The mean free path is the third property to consider. The mean free path of a
gas is the average distance between collisions of the gas particles and can thus
be used as an indication for the reaction rate. A short mean free path means
that the gas particles are more likely to collide and therefore more reaction will
occur within the gas. Evidence of this is the increase in cluster size with heavier
carrier gasses, as described above for different gas species [Chap.4.1.1.]. Also,
the mean free path can be derived from measurements on the state of the gas, because it relates to the
temperature and the pressure of the process, given by formula 4.1. In with λ is the mean free path; R the gas
constant; T the temperature; d the particle radius; N Avogadro’s number and P the pressure. From a research
point of view this relation can be very interesting for the choice in variable variation during experimentation.
4.1.4. Dwell time
The last gas property under consideration is the dwell time of a cluster. The dwell time is the time a cluster
needs to travels from sputtering of the target to the exit orifice at the end of the aggregation volume. It is the
most important factor with respect to controlling the final size and size distribution of the deposited clusters.
Evidently, the clusters can grow bigger if they travel through the growth region longer. This will also increase
the spread in the size distribution [8]. As mentioned before, if the dwelling time is too long, clusters can also
undergo possible undesirable agglomeration [Chap.3.2.].
4.2. System variables
Influencing the four properties can be done by changing the variables of the system. The complexity lies in the
fact that many gas properties can be affected at once by changing a single variable. For example, increasing the
argon flow results in a higher partial argon pressure and hence higher sputtering rate, but will also result in a
smaller mean free path because of collision cooling and shorter dwell time by increased flow. Another
complicating factor is that all variables have a limited operating window in which variations are noticeable.
Formula 4.1: mean free path
9 / 38
Cooling for instance, plays an important role in the overall process, but experimentation at various
temperatures will show no change if the cooling rate is dominated by the sweeping effect of a high gas flow.
Furthermore, some system variables prohibit the occurrence of the sputtering process at all if they cannot or
are not set within their functional operating window.
4.2.1. Partial pressures
Starting at the beginning of the aggregation volume, the first variables of the process are the flows of the
carrier gasses argon and helium supplied to the system. Previous research shows that a high argon pressure
generated by a high argon flow, has a linear correlation to the size of copper clusters. The higher the flow, the
bigger the clusters, because of an increase in the sputtering gas. However, a maximum will be reached. If the
flow is too high, the sweeping effect of the argon flow will decrease the dwell time of the clusters in the
aggregation volume and thereby decrease the mean cluster size and distribution [13, 15, 21].
For the partial helium pressure the correlation to the cluster size is opposite to that of argon. Increasing the
helium admixture of the total carrier gas will decrease the cluster size [1, 13, 15]. From this it can be concluded
that the cooling effect by helium’s high thermal conductivity is weaker than its sweeping effect, since the
former would encourage cluster growth and the latter prevents it. This does not mean however that helium’s
cooling effect is useless as its heat transfer aids the confinement of the target material vapor to create a
supersaturated gas state. This allows for a lower argon flow while maintaining supersaturation.
As for the total pressure, increasing the total gas flow will give more mass input which will proportionally
increase the pressure because of the fixed size exit orifice. An increase in available gas will have a positive
effect on the sputtering process, but also decrease the dwell time because of the higher flows. Thereby
increasing the production rate but decreasing the mean cluster size, if operated within the functional operation
window [14].
Besides changing the total pressure by varying the gas flows, the geometry of the aggregation volume can also
be changed to do so, for example by resizing the orifice opening. In theory, a smaller orifice will increase the
pressure in the aggregation volume and so supersaturation will be reached more easily. At the same settings
with respect to a normal orifice opening, this will increases the average cluster size because of a shorter mean
free path.
4.2.2. Magnetron strength and power
Increasing the target vapor yield can not only be done by
supplying more argon gas for a higher sputtering gas pressure, but
also by increasing the sputtering power to the anode. A higher
power supply will ionize more primary plasma. This generates
more particles for clusters growth, which results in a bigger
average cluster size [13, 15]. Also, research has been done on
different magnetron strengths. It was concluded that a weaker
magnetron results in smaller particles, as a stronger magnetron
leads to bigger particles. This implies that the sputtering yield can
be increased by a stronger magnetic field [13], which corresponds
to the confining effect of the magnetron head as described [see
Chap.2.1.].
4.2.3. Temperature
As described before, the temperature mainly influences the conditions for supersaturation. Decreasing the
temperature will drain more of the thermal and kinetic energy from the clusters and aid to achieve
supersaturation conditions more easily, therefore promoting cluster growth at equal settings [13]. Also, the
supersaturation can be limited by limiting cooling. By allowing nucleation to occur at relatively low
Figure 4.1: magnetron head without target material disk and anode.
10 / 38
supersaturation, the rate of particle formation is reduced, leading to lower concentrations, which results in
producing a much narrower size distribution [14].
4.2.4. Aggregation length
As for all possible variables of the system, the ability to change the length of the aggregation volume has the
biggest impact on the end result. By moving the magnetron head more to the front or the back of the
aggregation volume, the effective volume can be significantly changed. By increasing the distance to the orifice,
the dwell time of the clusters inside the volume is also increased. As the latter is key for their opportunity to
grow, this will mean that elongating the aggregation volume will strongly increase the mean cluster size and
distribution [15].
4.2.5. Bias voltage
Although it does not influence the sputtering or growth processes,
the bias voltages, which can be applied to the substrate holder,
can affect the clusters deposition. As mentioned above, a large
percentage of the clusters is ionized [Chap.4.1.1.], which allows
for acceleration of clusters exiting the orifice and subsequent high
energetic impact on a substrate. This process is called Energetic
Cluster Impact (ECI) and can be used for high quality thin-film
formation as well as smoothing rough surfaces. Typical energies of
5-15 eV per atom have been reported, which corresponds to
several km/s. Langevin molecular dynamics simulations show that
the impact zone of a cluster can be heavily deformed by pressures
up to 100 GPa with temperatures up to several 1000 K for a time
span of 10 ps. This process can completely disintegrate the
impacting clusters, accompanied by a diffusive motion of the
released atoms [6, 22].
4.3. Simulations and magic numbers
As cluster formation [Chap.3.] and effects of system variables cannot be studied inside the aggregation volume
during operation, but only deduced from experimental results, other ways to confirm the theory are required.
The most convenient and inexpensive method is by computer simulations and has been done in prior research.
Both Classical Nucleation Theory (CNT) simulations [Chap.3.1.], altered to incorporate both nucleus formation
and cluster growth [23], and Monte Carlo simulations [17] based on the previously described theory, yield
results comparable to those produced by real experiments. Both approaches show similar dependencies of the
system variables to the gas properties and result in the comparable log-normal size distributions, thus
supporting the validity of the process description.
Also, both simulations and experimental results suggest that there is a specific set
of ‘magic numbers’ with which clusters prefer to exist. These magic numbers
stand for the favorable total amount of atoms which are needed for a stable
cluster and correspond to complete atom shells. The reason magic numbers exist
is because several cluster parameters, like: the specific binding energy of atoms,
ionization potentials, cluster affinity to electrons and other parameters as a
function of atoms count, have energetic favorable extrema at magic numbers
[24]. In principle this directly implies that CNT does not hold, because then the
free energy of clusters as a function of their size is a smooth function and with
magic numbers the lowest energy path obtains a saw-tooth shape. Moreover, it
also indicates that various paths are possible in the complex energy landscape.
For geometric shell clusters based on twelve-vertex polyhedral (e.g. icosahedra,
Figure 4.2: ECI simulations at (a) 1.25 kV (b) 2.5 kV, (c) 5 kV, (d) 10 kV [23].
Figure 4.3: 13 copper atoms in cuboctahedron structure [3].
11 / 38
decahedra and cuboctahedra) like copper, these can be calculated by the formula:
[16] Giving the set of magic numbers: N = 13, 55, 147, 309, 561, … . Here, K stands for the
concentric shell number. So for a stable cluster consisting of only a single copper atom core with one shell (K =
1), the formula gives a total amount of 13 atoms in a twelve-vertex polyhedral structure, like figure 4.3. Since
calculations on the geometrical and electrical structure on small (magic number) copper nanoclusters yield the
same results as experimental data [3], it can be concluded that the theory holds. Even more conclusive results
stems from mass-spectroscopy on inert gas clusters, as the mass spectrum shows clear peaks at magic number
atom counts, thereby proofing the correctness of the theory [24].
5. Experimental setup
5.1. Equipment
For the production of copper nanoclusters a Nanosys550 Deposition System from MANTIS Deposition LTD was
employed. The Nanosys550 is equipped with a Nanogen 50 nanoparticle sputtering source. The Nanogen 50
consists of a magnetron head with a 1 Tesla magnetron, on which a 2 inch magnetron sputter source, i.e. disk
of target material, can be fixed. The target material was a Copper Sputtering Target from Alfa Aesar, 50.8mm
(2.0in) diameter x 3.18mm (0.125in) thick, purity 99.999% (metals basis). On top of the target material a
cylindrical anode is mounted, connected to a TDK-Lambda Genesys Gen600-1.3 programmable DC power
supply with a range up to 600 V and 1.3 A. During deposition the power was controlled by a build-in voltage
and current limiter.
The system is cooled by water flow from an external
heat sink system with a temperature of roughly 15
degrees Celsius. It was connected in series to several
components of the system: first, to the magnetron
head to reduce the resistive heating of the magnetron,
second, to the aggregation volume wall to cool the
sputtering process and third, to the two turbo vacuum
pumps.
The aggregation volume has a diameter of 13 cm. The
length of the aggregation volume can be varied from
approx. 8.5 cm to 18.5 cm by moving the magnetron
head to the front and back of the volume with a linear
actuator. The orifice between the aggregation volume
and deposition volume, has a diameter of 5 mm.
As carrier gasses argon and helium were chosen due to easy ionization and high thermal conductivity,
respectively. Both were supplied to the aggregation volume through separate channels of a MANTIS MFC (Mass
Flow Controller), each adjustable between 0 to 100 sccm (standard cubic centimeter). The purities of the main
elements argon and helium in the gas cylinders from Linde Gas Benelux B.V. were coded as 4.6 for 99.996 %
and 5.0 for 99.999 %. For the Half Factorial Design (HFD) experiments both argon and helium were of grade 4.6
[Chap.5.2.]. For the Energetic Cluster Impact (ECI) experiments argon of grade 5.0 was also used [Chap.5.3.].
Argon has a molecular weight of 39.95 kg·kmol−1
, first ionization of 1520.6 kJ·mol−1
, thermal conductivity of
17.72 x 10-3
W·m−1
·K−1
and viscosity 22.9 µPa·s (at 100 kPa, 300 K). Helium has a molecular weight of 4.00
kg·kmol−1
, first ionization of 2372.3 kJ·mol−1
, thermal conductivity of 0.1513 W·m−1
·K−1
and viscosity 20.0 µPa·s
(at 100 kPa, 300 K).
Initially, for both the HFD and ECI experiments, the carrier gasses were supplied to the system through 6 mm (1
mm wall) blue nylon tubing. During the ECI experiments oxygen intrusion was designated as a source of
Figure 5.1: experimental setup, affectionately called ‘Spoetnik’.
12 / 38
interference. Because of this, Varian CP17970 moisture and CP17971 oxygen washers were installed for the last
three ECI experiments. Simultaneously, the nylon tubing was changed to 1/4 inch (1/8 inch wall) stainless steel
tubing, to resolve the intrusion problems [see Chap.7.3.].
Before experiments, the aggregation volume was regularly baked by wrapped around metal braided heating
tape, which was later automated by a HTC-5500-pro heating controller. Baking the system gave a base pressure
before experimentation of 10-8
mbar (UHV, Ultra High Vacuum) and operating pressures of 10-3
mbar (HV, High
Vacuum) in the deposition volume, with 10-1
mbar (MV, Medium Vacuum) at the gas input of the aggregation
volume. Baking is a cleaning necessity in vacuum systems to evaporate unwanted substances of the wall of the
aggregation volume, thereby preventing degassing of these substances during vacuum pumping, which would
prevent reaching UHV. The pressure in the deposition volume is created and maintained by a 300 litres/second
Leybold Turbovac SL 300 pump. Differential pumping of the Nanogen 50 is performed by a second Leybold
Turbovac SL 300 pump. Both pumps are backed by a single Varian SH110 dry backing pump.
During all experiments two substrates were placed in the substrate holder,
a SimPore Silicon Nitride (9 window,5 nm thickness) TEM grid and a cut
piece (roughly 1 cm2) of polished P/Boron <100> silicon wafer [figure 5.2].
To control the deposition time on the substrates, the substrate table could
be blocked from the cluster flow by a shutter plate. In the second
experiment series on Energetic Cluster Impact (ECI), the substrate holder
was connected to a positive bias voltage up to 6 kV by a MANTIS STC
(Substrate Table Controller). Although not used during any experiment,
the substrate holder could also be rotated up to 20 rpm.
To gain more insight in the yield during operation, a custom build Quartz
Crystal Microbalance (QCM) connected to an Agilent 53220A frequency
counter [25] was placed in the deposition volume at the edge of the
cluster flow. Even though no quantitative measurements were done with
this QCM it was still important to gain a qualitative insight in the
sputtering process.
Imaging of the TEM grids for both HFD and ECI experiments, was done using a JEOL 2010F Transmission
Electron Microscope (TEM). Images were recorded on a CCD camera at nominal microscope magnifications of
200k, 300k, 400k and 500k. The functioning and imaging importance of the TEM is explained in more detail in
the results chapter [Chap.6.1.1.]. For the ECI depositions additional analyses was performed using a Veeco
Multimode PicoForce Atomic Force Microscope (AFM). The AFM was controlled by a Nanoscope V controller
and Nanoscope v7.3 software, operating in tapping mode at atmospheric pressure. The first AFM images where
made with a ‘sharp’ tip model PPP-NCHR from NanoAndMore GmbH, with a tip radius of 7 nm. Due to
unexpected tip problems and low image detail the tip was changed to a ‘supersharp’ tip model SSS-NCHR from
NanosensorsTM
, with a tip radius of 2 nm. The functioning and imaging importance of the AFM is also explained
in more detail in the results chapter [Chap.6.1.2.].
For the HFD experiments, size and distribution calculations of the produced particles were done with imaging
software Image-Pro® Plus 7 from Media Cybernetics. For the final analyses of the HFD results statistical
software package Minitab® 16 was used.
5.2. Settings: Half Factorial Design
In the first experiment series, the operating window of the Mantis Nanosys550 Deposition System was
characterized by a Half Factorial Design (HFD) [see Appendix A]. For this HFD four factors were chosen: (a)
argon flow, (b) helium flow, (c) sputtering power and (d) aggregation length. The choice for these factors was
relatively straight forward because they are known to be the key factors in the sputtering process and
therefore the main controls on the Mantis system. After literature research, it was decided not to use the
Figure 5.2: substrate holder with TEM grid in the middle and silicon wafer piece in the top window.
13 / 38
center coordinates approach, as described for the (hyper-)cubes in a factorial design, but to go with settings of
similar systems from equivalent research found in literature.
The input of the two gasses was precisely set on the Mass Flow Controller. The +, 0 and - settings for argon
were chosen to be 70 (high), 60 (center) and 50 (low) standard cubic centimeter per minute (sccm). For helium
20 (high), 10 (center) and 0 sccm (low) were chosen. For controlling the sputtering power the voltage of the DC
Power Supply was manually maintained around 300 V. The current was set to 0.3 (high), 0.2 (center) and 0.1 A
(low) to yield the desired power of approximately 90, 60 and 30 W. To decide on the several settings for the
aggregation length an arbitrary length scale had been defined on the telescopic hull of the linear actuator
between two points on its top and bottom. Between these two the aggregation lengths of 10 (high), 7.5
(center) and 5 cm (low) were used as the fourth and therefore confounding factor d.
To research the errors and stability of the system six Center Point Experiments (CPEs) [Appendix A.3] were
incorporated in the experimental design. From these six, two were placed at the beginning and two at the end
of the experiment series. A fifth was placed after the first experiment and the sixth in the middle of the series.
Together all settings comprised the experimentations as shown in table 5. The pressure values in the last
column were measured inside the deposition volume and were denoted in order to monitor if the sputtering
system was in right working order.
Table 5.: Experimental setup by Half Factorial Design with Center Point Experiments
Experiment # Ar flow
(sccm)
He flow
(sccm)
Length
(mm)
Voltage
(V)
Current
(A)
Power
(W)
Pressure
(mbar)
0 60 10 75 320 0,200 64 1.2 .10-3
0 60 10 75 329 0,200 66 1.2 .10-3
1 50 0 50 290 0,100 29 9.0 .10-4
0 60 10 75 296 0,200 59 1.2 .10-3
2 70 0 100 280 0,100 28 1.3 .10-3
3 50 20 100 300 0,100 30 1.0 .10-3
4 70 20 50 293 0,100 29 1.4 .10-3
0 60 10 75 316 0,200 63 1.2 .10-3
5 50 0 100 315 0,300 95 9.1 .10-4
6 70 0 50 305 0,300 92 1.3 .10-3
7 50 20 50 330 0,300 99 9.6 .10-4
8 70 20 100 300 0,300 90 1.4 .10-3
0 60 10 75 296 0,200 59 1.2 .10-3
0 60 10 75 293 0,200 59 1.2 .10-3
To make sure that the system was operating in a steady state it was
turned to the HFD specified settings and run for 10 minutes before
experimentation. During this time the shutter plate in front of the
substrates was closed, see figure 5.3. This allowed the process variables
(e.g. cooling and vacuum) and cluster growth reactions to settle their
operating equilibria. After this equalization time, the shutter plate was
opened for 10 minutes to deposit the aggregated clusters on the TEM grid
and silicon wafer substrates.
These experimentations were performed by supervisor Ten Brink, who
generously supplied the required information and resulting data for this
thesis [Acknowledgments].
Figure 5.3: substrate holder in deposition volume with closed shutter plate.
14 / 38
5.3. Settings: Energetic Cluster Impact
In the second experiment series, research was done on the acceleration of copper clusters for Energetic Cluster
Impact (ECI). The name ECI was proposed by Haberland et al. [6] As described before, the majority of the
copper clusters are expected to have a net negative charge [Chap.4.1.1.], which can be utilized for acceleration
experiments by applying a positive bias voltage to the substrate holder [Chap.4.2.5]. Experiments where
preformed with high and low bias voltages at several system settings. For the settings of this experiment series
no specific experimental design was employed, in contrast to the HFD experiments. The reason for this is that
analyses of the HFD experiments showed that the system suffered from, at that time, unknown interference
which influenced the sputtering yield [see Chap.7.3.]. The settings used for the ECI depositions were chosen in
such a way that they could potentially simultaneously reveal the true nature of the discovered interference as
well as produce comparable data for ECI. Table 6 shows the settings for the performed experiments, with
paired high and low bias voltages at equal settings. The settings for the aggregation length and the sputtering
power were set and controlled similar to the HFD experiments. Further operating procedures, like the 10
minute equalization and deposition times, were also maintained [see Chap.5.2]. For the first two experiments
the pressure inside the deposition volume during operation was not logged, unfortunately.
As oxygen intrusion was designated the prime suspect for causing the experienced interference, one of the
changes made to the system was the upgrade of the argon sputtering gas purity. This is why not only the gas
flows are noted, but also the gas purities as coded by the supplier: [4.6] for 99.996 % and [5.0] for 99.999 %. In
a further attempt to reduce the suspected oxygen intrusion moisture and oxygen filters were installed for the
last three ECI experiments. The filters were placed in series on the argon gas cylinders output at the beginning
of the supply tubing to the system. Simultaneously, the nylon tubing was replaced by stainless steel to
eliminate suspected gas diffusion through the tube wall.
Table 6.: Experimental setup for Energetic Cluster Impact experiments
Experiment # Ar flow (sccm) [Purity]
He flow (sccm) [Purity]
Length (mm)
Power (W)
Bias voltage (kV)
Pressure (mbar)
1 60 [4.6] 10 [4.6] 75 60 1 - 2 60 [4.6] 10 [4.6] 75 60 6 -
3 70 [4.6] 20 [4.6] 50 90 0 1.4 .10-3
4 70 [4.6] 20 [4.6] 50 90 6 1.4 .10
-3
5 70 [5.0] 20 [4.6] 50 96 3 1.4 .10-3
6 70 [5.0] 20 [4.6] 50 93 6 1.4 .10
-3
7* 70 [5.0] 20 [4.6] 50 90 6 1.4 .10-3
8* 70 [5.0] 20 [4.6] 50 90 -> 69 ** 0 1.4 .10
-3
9* 70 [5.0] 0 50 81 6 1.3 .10-3
* Experiments preformed after installation of filters an stainless steel tubing ** Constant power drop during experiment
6. Results
The following paragraphs show the obtained results from the Half Factorial Design (HFD) and the Energetic
Cluster Impact (ECI) experiment series together with a description of the used analyses techniques. First a short
description of the imaging techniques is given to reveal the important features of their images.
6.1. Imaging
6.1.1. Transmission Electron Microscope
In a conventional Transmission Electron Microscope (TEM) a specimen is irradiated by an accelerated electron
beam of uniform current, which passes through the specimen. In order for the electron beam to pass, the
15 / 38
specimen must be very thin, typically between tens to hundreds nm.
For a thicker specimen, the passing electron beam will be scattered
too much and the specimen internal structure information as
collected by the passing electrons, will be lost. For the JEOL 2010F
[figure 6.1 (a)], employed during this research, the electron beam is
created by a field emission gun in the top of the TEM vacuum column
and accelerated by a 200 kV acceleration voltage. A staged
condenser-lens and aperture system underneath the gun permits
variation of the illumination aperture and the area of the specimen
illuminated. Below the specimen a second lens and aperture system
regulates the beam intensity distribution of the passed electrons and
focusses them onto a fluorescent screen. The created image can be
recorded by a CCD camera, attached to the wide-view port above the
fluorescent screen. The contrast mechanisms of a TEM are quite
complex, therefore they will be explained in a simplified manner that
is adequate in the context of the present work. The local contrast of
the image arises from two specimen properties. The first is
comparable to a normal light microscope, as thick areas of the
specimen will be dark in the resulting images. In TEM imaging the
thickness perceived by the passing electrons results from the
numbers and weight of atoms present in their path. This so-called
mass-thickness contrast means that a specimen of uniform thickness
will show bright and dark regions in the image for corresponding
parts of the specimen containing low and high mass-density,
respectively. If the specimen has uniform density, then the thicker
regions will appear darker in the image. This contrast mechanism
results from the number of charges that an encountered atom
carries. The likelihood that an electron is deviated from its direct path
by an interaction with an atom increases with the number of charges.
Therefore, heavier elements represent more powerful scattering
centers than light elements. Due to this increase of the Coulomb
force with increasing atomic number, the contrast of areas in which
heavy atoms are localized will appear darker than of such comprising
light atoms. If a substrate is used for holding the specimen material
(e.g. for deposition of clusters) and the internal atomic structure of
that substrate is known exactly, it can be subtracted from the TEM
image, resulting in a clear representation of the deposited material.
The interaction of the passing electrons with the specimen atoms can
result in a very direct and comprehensible image with atomic
resolution. This means that the TEM image gives a natural 2-
dimensional representation of the specimen, comparable to normal
light microscopy, as can be seen in figure 6.1 (b) [26].
The second contrast mechanism is called diffraction contrast. When
the electron beam of the TEM interacts with a crystal in the sample
(which is sufficiently thin) diffraction due to elastic scattering will
occur, resulting in diffracted (hkl) beams next to the transmitted
(000) beam. Depending on the crystal orientation more electron
intensity can be scattered into the diffracted beams compared to the
000 beam. For standard conventional imaging generally a relatively
small objective aperture is used centered around the 000 beam. This
(a) JEOL 2010F TEM
(b) Example image (ECI exp.3)
Figure 6.1: TEM equipment and image
(a) Veeco Multimode PicoForce AFM
(b) 2D example image (ECI exp.8)
(c) 3D example image (ECI exp.8)
Figure 6.2: AFM equipment and image
16 / 38
means that all diffracted beams are excluded for image formation. This type of imaging is called bright field
imaging. Some crystals will appear dark when they strongly diffract, e.g. when the incident electron beam is
nearly parallel to a so-called (low index) zone axis of the imaged crystal. This holds for several particularly black
clusters (with a nearly cubic shape) in the image depicted in figure 6.1. So, like figure 6.1 all the TEM imaged we
recorded were bright-field images showing predominantly diffraction contrast, but also significant mass-
thickness contrast.
6.1.2. Atomic Force Microscope
An Atomic Force Microscope (AFM) [figure 6.2 (a)] uses a sharp tip mounted on a cantilever spring to scan the
surface of a specimen. The cantilever is typically silicon or silicon nitride with a tip radius of curvature on the
order of nanometers. The image is constructed by measuring the angle of the cantilever, which changes
because of the changing force between the specimen and the tip, as it moves over the specimen surface. The
latter is done by a laser reflecting of the back of the tip onto a photodiode. A feedback loop reacts to the
change in laser position by adjusting the height of the sample to keep the force between the surface and the
tip constant. Depending on the situation, forces that are measured in AFM include mechanical contact
force, Van der Waals forces, capillary forces, chemical bonding, electrostatic forces, magnetic forces, Casimir
forces, solvation forces, etc. The surface topography image acquired by the AFM consists of the height
adjustments made by the feedback loop and therefore represents the height profile of the scanned specimen.
By adjusting the feedback controls, the deflection of the tip and the related height adjustment of the specimen
can be minimized, which increases the accuracy of the AFM image. For soft and fragile specimens a dynamic
operation mode called ‘tapping’ is used to prevent possible damage to the specimen surface by the applied
forces (in the nN range for contact mode). As the name implies the tip oscillates up and down during scanning.
In tapping mode the amplitude modulation, at resonance frequency oscillation, generated by the interaction
forces of the tip and surface is used as the feedback signal for constructing the image. In contrast to TEM
imaging, AFM generates a less direct representation of the specimen surface, because of the utilized
interactions between the surface and tip and the strong dependency on operation settings, like scanning
speed. The advantage of AFM over TEM is that AFM imaging records the height profile, thereby generating 3-
dimensional information of the specimen surface. Figure 6.2 (b) shows an 2-dimensional example image with
scale bar for height. Figure 6.2 (c) shows a 3-dimensional representation of figure 6.2 (b). Also, AFM imaging is
easier to perform with respect to TEM imaging. Its relatively simple operating and versatility make it ideal for
fast examination of specimens, but caution has to be exercised for the mentioned probability of damage and
strong dependency on operation settings for the imaging quality. During this research the AFM was operated in
tapping mode to prevent contact with the deposited clusters [27].
6.2. Results: Half Factorial Design
Although the clusters where deposited on both a TEM grid and a piece of silicon wafer, only the TEM grids
were used for analyses of the HFD experiments. The silicon wafer pieces were stored in the anticipation that
other methods of analyses, e.g. AFM, could be required. Imaging of the TEM grids was done with a JEOL 2010F
TEM. Two separate series of images were made at the magnifications 200k, 300k, 400k and 500k (i.e. 500.000),
comprising a total of 8 images per grid. From these images the 300k and 400k were chosen for quantitative
analyses on the amount and size of the deposited particles.
6.2.1. Image Quantification Analyses
Quantification of the deposited clusters was done by the image analysis software Image-Pro® Plus. Before this
could be done the images had to be ‘cleaned’ by applying several software filters. First a threshold was set for
flattening all shades of gray of the TEM grid background. This significantly reduced the image noise and clearly
revealed the deposited particles as dark objects. Second, the software settings for counting objects were
calibrated to count dark objects of a certain area and diameter. This was done manually and checked visually to
ensure that the software recognized only real particles. Third, a Watershed Split algorithm was applied to the
image to split agglomerated clusters from each other. This algorithm treats the gray levels of an image as the
17 / 38
topographical relief of the imaged objects, where every gray shade corresponds to a certain altitude. By doing
so, local minima can be recognized as a transition region between two particles, even if they lie against one
another. An example of the results can be seen in figure 6.3.
Figure 6.3: particle recognition (red borders) and counting (green numbers) by image analysis software (HFD exp.8).
After these alterations the software calculated the amount and size of the objects on the cleaned image. To
ensure that these values were valid, several of the results were checked by counting them manually. It was
concluded that the data extraction from the images was done correctly and that the values corresponded to
the physically deposited clusters. Since two series were made at both 300k and 400k magnification, the results
of these series were averaged to a final result. The results are shown in table 7, in which size denotes the
average diameter of the particles and SD the standard deviation of the size distribution. The 400k magnification
images, which contain a smaller amount of particles then the 300k images, showed an average of 107 clusters,
with a minimum of 21 clusters and maximum of 274 clusters.
Table 7.: Results of experiments on cluster size from images of 300k and 400k magnification.
300k 300k 400k 400k 300k 400k
Exp. # Size SD Size SD Size SD Size SD Size SD Size SD
1st
1st
2th 2th 1st 1st
2th 2th
0 10,10 2,85 10,70 3,20 9,79 2,48 10,00 2,71 10,4 3,0 9,9 2,6
0 11,89 4,53 12,96 4,87 10,73 4,52 12,30 5,31 12,4 4,7 11,5 4,9
1 13,41 4,76 12,40 3,74 11,88 5,62 10,60 4,30 12,9 4,3 11,2 5,0
0 13,41 4,76 12,40 3,74 11,88 5,62 10,60 4,30 12,9 4,3 11,2 5,0
2 22,40 6,93 24,18 7,35 20,29 7,79 20,80 10,11 23,3 7,1 20,5 9,0
3 15,75 4,02 15,53 4,19 14,87 2,69 12,96 5,20 15,6 4,1 13,9 3,9
4 12,31 4,68 12,44 4,22 11,24 5,12 10,89 5,18 12,4 4,5 11,1 5,2
0 15,91 4,38 16,07 5,24 13,91 5,82 14,86 4,91 16,0 4,8 14,4 5,4
5 16,47 7,76 15,26 4,37 15,24 7,75 14,44 5,27 15,9 6,1 14,8 6,5
6 14,56 5,37 14,27 5,20 13,36 5,53 11,75 6,04 14,4 5,3 12,6 5,8
7 12,04 4,76 12,17 5,14 10,47 4,43 10,38 4,64 12,1 5,0 10,4 4,5
8 16,02 4,18 16,54 3,64 16,51 3,58 16,39 3,44 16,3 3,9 16,5 3,5
0 18,68 5,51 18,41 5,16 16,65 5,43 16,09 4,92 18,5 5,3 16,4 5,2
0 19,33 6,06 19,08 5,41 17,86 4,68 17,57 6,57 19,2 5,7 17,7 5,6
18 / 38
Table 8.: Factorial Design analysis by Minitab 16.
300k 400k Term Effect Coefficient Effect Coefficient
Constant (Average) 15,36 13,88 Argon 2,46 1,23 2,55 1,27
Helium -2,52 -1,26 -1,83 -0,92 Current -1,39 -0,69 -0,62 -0,31
Aggr.Length 4,82 2,41 5,12 2,56
Center Points 0,58 0,38 S = 15,8995 S = 14,3672
6.2.2. Statistical Data Analyses
The results in the last four columns of table 7 together with the settings from table 5 comprise the data needed
to analyze the aliases from the HFD experiments [Appendix A]. This final analysis was performed by the
statistical software package Minitab® 16. The program was instructed to handle the data as a four factor HFD
with six center point experiments. Table 5 was imported as the high/low factor settings and table 7 as the
associated aliases. For an extra accurate analysis, the standard deviations in table 7 were squared and also
imported as the variances of the average diameters at 300k and 400k. The results of Minitab are shown in table
8. Here S denotes the estimated standard deviation of the error in the model, calculated from the variances.
The ‘Effects’ are the calculated size increases (in nm) between the high and low settings, resulting in
‘Coefficients’ half their size (in nm). Figures 6.4 and 6.5 show the linear regression plots of these effects.
10-1
17
16
15
14
13
10-1
10-1
17
16
15
14
13
10-1
Argon
HFD
esti
ma
ted
dia
me
ter
Helium
Current Aggregation Length
Response
Center Point
Point Type
Main Effects Plot for 300k images
Figure 6.4: Estimated effects for cluster diameters, calculated from the 300k images with the HFD. All effects are plotted with respect to the estimated Constant (Average).
19 / 38
10-1
16
15
14
13
12
10-1
10-1
16
15
14
13
12
10-1
ArgonH
FD e
sti
ma
ted
dia
me
ter
Helium
Current Aggregation Length
Response
Center Point
Point Type
Main Effects Plot for 400k images
Figure 6.5: Estimated effects for cluster diameters, calculated from the 400k images with the HFD. All effects are plotted with respect to the estimated Constant (Average).
6.3. Results: Energetic Cluster Impact
For the analyses of the ECI experiments, on cluster acceleration by an applied bias voltage, both the TEM grids
and the silicon wafer pieces were used. Imaging of the TEM grids was performed with a JEOL 2010F TEM by
taking two image series at several magnifications, similar to those of the HFD experiments [Chap.6.2.]. The
silicon wafer pieces were mapped using a Veeco Multimode PicoForce AFM. The reason not to use TEM grids
for AFM imaging at first was to avoid potential damage to the TEM grids, which were considered more valuable
as TEM is a more direct imaging technique. Later on AFM was also done on TEM grids to examine imaging
technique differences. For equal comparison of all ECI experiments, the 300k magnification is mainly shown. At
this magnification most particles for yield comparison are shown, while still maintaining sufficient cluster shape
details. In the discussion of the ECI results [Chap.7.2.], other magnifications were also used for more detail.
6.3.1. TEM images of high/low bias voltages
TEM images were recorded for analyses of bias voltage effects on the yield and shape of the deposited clusters.
Figure 6.6 shows the paired TEM images from equal deposition settings with low (left) and high (right) bias
voltages. A summary of the used settings [from Chap.5.3., table 6] is noted underneath the images for more
easy comparison purposes [see Chap.7.3.]. These ECI TEM images were not analyzed or quantified by software,
as was done with the HFD TEM images, because of the clearly visible but unexpected results. Only one
experiment was performed with the settings of experiment 9 because both the Quartz Crystal Microbalance
(QCM) frequency during deposition and TEM images afterwards showed an unsuccessful deposition.
20 / 38
(a) ECI Exp.1: 1 kV bias (b) ECI Exp.2: 6 kV bias Settings(1): 60 sccm Ar[4.6] + 10 sccm He[4.6]
@ 75 mm and 60 W
(c) ECI Exp.3: 0 kV bias (d) ECI Exp.4: 6 kV Settings(2): 70 sccm Ar[4.6] + 20 sccm He[4.6]
@ 50mm and 90 W
(e) ECI Exp.5: 3 kV bias (f) ECI Exp.6: 6 kV bias Settings(3): 70 sccm Ar[5.0] + 20 sccm He[4.6]
@ 50mm and 96 W(left) and 93 W(right)
50 nm
21 / 38
(g) ECI Exp.8: 0 kV bias (h) ECI Exp.7: 6 kV bias Settings(4): 70 sccm Ar[5.0, filtered & steel tubing] + 20 sccm He[4.6, steel tubing]
@ 50 mm and 90 W(left) and 90->69 W(right)
(i) ECI Exp.9: 6 kV bias
Settings(5): 70sccm Ar[5.0, filtered & steel tubing] @ 50 mm and 81 W
Figure 6.6: TEM images of TEM grids at low (left) and high (right) bias voltages.
6.3.2. TEM and AFM images
Because TEM imaging results in a 2-dimensional representation of the thin membrane substrate containing
particles, additional topographic information was acquired by AFM imaging. The height information of the AFM
images can reveal if the deposited clusters had been deformed, i.e. flattened or disintegrated, due to energetic
impact or maybe even imbedded into the substrate material. Figure 6.7 shows TEM images (left) and AFM
images (right) of the same experiments side-by-side. To investigate the validity of the comparison between
TEM and AFM imaging, with regard to the two different substrates, AFM images were made of a TEM grid
[figure 6.7 (f)] and silicon wafer pieces [figure 6.7 (b) and (d)], as noted underneath the images.
During AFM imaging problems with the sharp tip were encountered, leading to seemingly unrealistic and low
detail images. To avoid these problems the AFM tip was changed from a sharp to a supersharp tip. Figure 6.8
shows the acquired AFM images made by these two tips next to a TEM image, all from ECI experiment 8. Both
tips were used on the silicon wafer piece for tip comparison [figure 6.8 (b) and (d)]. Additionally, the
supersharp tip was used for imaging of the TEM grid for comparison to the silicon wafer piece [figure 6.8 (c)
and (d)] and comparison to the TEM image [figure 6.8 (a) and (c)].
50 nm 50 nm
22 / 38
(a) ECI Exp.2: TEM grid
(b) ECI Exp.2: silicon wafer, 1 by 1 um (sharp tip)
(c) ECI Exp.3: TEM grid
(d) ECI Exp.3: silicon wafer, 1 by 1 um (sharp tip)
(e) ECI Exp.4: TEM grid
(f) ECI Exp.4: TEM grid, 1 by 1 um (sharp tip)
Figure 6.7: TEM images (left) compared to AFM images (right) from the same experiments.
23 / 38
(a) ECI Exp.8: TEM grid (b) ECI Exp.8: silicon wafer, 5 by 5 um (sharp tip)
(c) ECI Exp.8: TEM grid, 2 by 2 um (supersharp tip)
(d) ECI Exp.8: silicon wafer, 2 by 2 um (supersharp tip)
Figure 6.8: TEM image (a) for comparison to AFM image with sharp tip (b) and supersharp tip (c, d) from ECI experiment 8.
7. Discussion
7.1. Discussion: Half Factorial Design
7.1.1. Factor effects
As can be seen from table 8 and the figures 6.4 and 6.5 [Chap.6.2.2.], both magnifications give comparable
results. The biggest noticeable difference is that all plotted values of the 300k images lie approximately 1.5 nm
higher than those of the 400k images. This results from the difference in average cluster size of 15.36 nm
(300k) and 13.88 nm (400k) relative to which all effects are plotted. An explanation for this difference can be
found in reviewing the image quantification analyses [Chap.6.2.1.]. After close inspection of the images it is
believed that as the magnification of the TEM grid surface is increased the round edges of the copper clusters
become more clearly visible. As a result, the difference between the grid surface and cluster edges becomes
less distinct, because of a more gradual transition by the round edge shape of the clusters. This results in the
recognition of a smaller cluster, as the boundary of the recognized particle by the software is placed on top of
the round edge instead of next to it, thereby lowering the average particle size at higher magnification. Still, for
future research it is suggested to use a similar magnification. A smaller magnification would result in a bigger
50 nm
24 / 38
spread in particle size, as the conversion factor of nanometers per pixel becomes bigger for the fixed resolution
of the TEM’s camera. This will cause a bigger spread, as the difference of a single pixel between particles will
translate in a bigger size difference.
With respect to the slopes of the main effects, it can be seen that most correspond to the described influences
of the system variables [Chap.4.2.] An increase in argon flow results in bigger clusters, where an increasing
helium flow decreases the cluster size. The change in aggregation length also shows the expected effect, as
particles can grow bigger for a longer aggregation length. Only the effect of increasing current is not as
expected. A higher current supplies more sputtering power and should increase cluster size according to
previous research. To see if this calculated effect and those of argon, helium and the aggregation length are
reliable, the significance of these effects has to be investigated.
7.1.2. Significance of factor effects
When calculating the average standard deviations for the 300k and 400k images from table 7 (3th and 1st
columns from the right) [Chap.6.2.1.], one gets 4.9 nm and 5.1 nm respectively. If these averages are compared
to the main effects from table 8 [Chap.6.2.2.], it can be seen that even the biggest main effect of 4.8 nm and
5.1 nm by the aggregation length is equal at best. The other main effects are even smaller. Also, the S-values
from table 8 show an extremely high standard deviation compared to all other values from the Half Factorial
Design (HFD). From this it can be concluded that the calculated effects from the HFD are not high enough to be
significant, since they easily fall within the error margin of the experiments.
From these comparisons a number of possible problematic scenarios arise. First of all it is possible that the
differences between the high and low settings from the HFD are too small in comparison to the natural spread
in the sputtering process for this specific system. Although this seems unlikely with respect to similar research
found in the literature [1]. Providing that the operating windows of the several factors allow a larger difference,
increasing the difference between the high and low settings will show larger effects. This can potentially yield
significant values for the characterized factors. Second, it is possible that the minimal amount of experiments
preformed during this characterization is insufficient to reveal the true effects of the system variables
[Appendix A.2.]. Completing the HFD to a FD and addition of replica experiments will increase the reliability of
the responses. Third, it can be that the system suffers from non-controllable influences or possesses ignored
characteristics, which have a strong effect on the cluster growth process. Fortunately, investigation of the
latter can be done by the CPEs [Appendix A.3].
When looking at table 9, which only contains the CPE results and their dates, it shows that there is some
accumulating effect as every subsequent CPE results in a bigger average cluster size and standard deviation.
When examining the size steps between subsequent average sizes, it looks plausible that the increasing cluster
size is primarily a consequence of the time passed between the CPEs, not the experiments in between. CPE 1
and 2 were performed in two consecutive days without HFD experiments and have average size differences of
2.0 nm (300k) and 1.6 nm (400k). Between CPE 3 and 4 three HFD experiments were performed in eleven days,
resulting in average size differences of 3.1 nm (300k) and 3.2 nm (400k). Between CPE 4 and 5 four HFD
experiments in two days were performed, but the average size differences only measured 2.5 nm (300k) and
2.0 nm (400k) [see also Chap.6.2.1., table 7].
Although these averages show that the performed HFD experiments do not seem to cause the difference, there
is also no simple correlation visible between time and size, only an increase. An obvious, but definitely not
trivial assumption is that the process is well reproducible. In the present research for instance the target
material is continuously eroded, thereby obtaining a deeper race track [Chap.2.1.]. This can potentially affect
the reproducibility of subsequent experiments in a negative way.
25 / 38
. Table 9.: Results of Center Point Experiments with dates
300k 400k
Experiment CPE # Experiment Date Size SD Size SD
0 1 November 17th 2010 10,4 3,0 9,9 2,6
0 2 November 18th 2010 12,4 4,7 11,5 4,9
1 HFD experiment
0 3 November 19th 2010 12,9 4,3 11,2 5,0
3 HFD experiments
0 4 December 1st 2010 16,0 4,8 14,4 5,4
4 HFD experiments
0 5 December 3rd 2010 18,5 5,3 16,4 5,2
0 6 December 3rd 2010 19,2 5,7 17,7 5,6
7.2. Discussion: Energetic Cluster Impact
As shown in the Energetic Cluster Impact (ECI) results chapter [Chap.6.3.] both TEM and AFM images were
recorded to reveal any changes due to the applied bias voltages to the substrate holder. In the following two
paragraphs the expected and observed differences will be examined to reveal if the clusters underwent ECI,
thereby proofing their charge as described in literature [6, 22].
7.2.1 Bias voltage effects
TEM images were made [see Chap.6.3.1] to gain comparable shape and yield data at different bias voltages. If
the copper clusters were accelerated because of their charge, as described in the chapter on gas properties and
system variables [Chap.4.1.1. and Chap.4.2.5.], shape changes are expected to show. As described, ECI will lead
to deformation or complete disintegration of the energetic impacting clusters. When comparing the left and
right images from figure 6.6. [Chap.6.3.1.], no significant shape differences are observed. For a better visual
comparison the highest magnification (500k) images of all ECI experiments were compared side-by-side, in a
similar fashion to figure 6.6. As an example, figure 7.1 show the 500k magnifications of ECI experiments 3 (a)
and 4 (b), which were performed with similar settings but different bias voltages. From the 500k image
comparison it was concluded that there is no visible shape change supporting ECI.
(a) ECI Exp.3: 0 kV bias
(b) ECI Exp.4: 6 kV bias
Figure 7.1: 500k magnification TEM imaging of high and low bias voltage experiments, for inspection on shape change due to ECI
Also (with the expert help of the group leader professor Kooi), some high resolution TEM images were made of
ECI experiment 2 (preformed at 6 kV high bias voltage) for extra close cluster inspection. Figure 7.2 (a) and (b)
show 2 clusters which were found representative for cluster shape. Unfortunately, this also did not reveal any
26 / 38
clear proof of cluster deformation. Still, it has to be noted that the existence of charged clusters cannot be
rejected based on these results, as TEM imaging provide no height information. Although not studied in more
detail, these high resolution images show that the clusters all seem to have a core-shell structure, which can be
seen from the dark center with lighter shade ring. Additionally, interval imaging of cluster 2 showed that the
clearly visible atomic structure of the cluster [figure 7.2 (b)] rearranged under the influence of the TEMs
electron beam energy [figure 7.2 (c)].
(a) ECI Exp.2: cluster 1
(b) ECI Exp.2: cluster 2
(c) ECI Exp.2: cluster 2
(restructuring by TEM beam)
Figure 7.2: high resolution TEM imaging for extra close inspection on shape change due to ECI
With respect to the yield of the ECI experiments, there are some visible differences between the high and low
bias voltage experiments. When comparing the images of figure 6.6 [Chap.6.3.1.], almost all high bias voltage
experiments show a higher yield with respect to their low bias counterparts, with the exception of ECI
experiment 1 versus 2. The former could indicate that extra (charged) clusters deviate from their initial path
and deposit on the substrate, because of the attracting forces by the bias voltage. Although, this could reveal
the presence of ECI, it was concluded not to be a definitive proof. No replica experiments were performed to
reconstruct the obtained results and calculate their significance. This was not executed as the system was
changed several times to eliminate the suspected severe problem of oxygen intrusion. These changes resulted
in different (decreasing) yields and even to the point that finally no deposition occurred during ECI experiment
9. Because of these large yield differences and the fact that the results of ECI experiments 1 and 2 are
contradictory to the rest, the data was found not to be reliable enough to allow any convincing conclusions on
yield differences and therefore not on the presence of ECI.
A potential reason that the effect of ECI is not observable can be related to a too small (e.g. single elementary)
charge compared to a too high weight (i.e. too many atoms containing) cluster. Therefore 6 kV can be still
much too small to provide any observable effect, because with a cluster containing 20000 atoms corresponds
to only 0.3 eV per atom. Moreover, the nanoparticles are accelerated only when they exit the orifice of the
nanocluster source, because the bias voltage is applied to the substrate holder. Due to the pressure difference
already a high speed (sonic) particle beam is created and therefore the effect of applied bias voltage to the
substrate holder is only an additional effect that can be relatively modest and thus unobservable.
Although the ECI results from figure 6.6 do not show any ECI effects, they do show more proof of the HFD
results. It can be clearly seen that the described influences of system variables on gas properties [Chap.4.], like
increase in gas supply and purity from settings(1) to settings(3), increase the sputtering yield. Also, the
deposited clusters seem to be smaller with a shorter aggregation length, when comparing settings(1) with
settings(2) and settings(3).
7.2.2 Combining TEM and AFM
Since all expected ECI results, i.e. deformation, disintegration or embedding, result in changes in the cluster
height, topographical information from AFM imaging was required. Therefore AFM imaging was done on some
27 / 38
of the substrates of the ECI experiments, as shown in
figure 6.7 [Chap.6.3.2.] When comparing the TEM
and AFM images it must be concluded that the
results obtained with both techniques are rather
dissimilar even distinctly different. There is a large
difference in the observed particle shape and density.
On the TEM images it can be seen that the clusters
are agglomerated and have relatively large empty
spaces between them. On the AFM images however,
no similar shaped agglomerates are visible and the
observed particles show almost no empty space.
Also, when resizing the images of ECI experiment 2 to
become of equal scale, as has been done in figure
7.3, it can be seen that the particles on the TEM
image are several times smaller than those on the
AFM image.
At first, this difference was attributed to the different
substrate types used for AFM (silicon wafer) and TEM
(TEM grid) imaging. To examine this difference the TEM grid of ECI experiment 4 was also used for AFM
imaging, as can be seen in figure 6.7 (f). Although, this image suffers from some additional problems resulting,
e.g. horizontal lining, it can be seen that the particles are of comparable size and shape. The latter shows that
AFM imaging on TEM grids or silicon wafer pieces gave equal results for the used AFM operation settings.
Because of the encountered problems of dissimilar results, the tip of the AFM was changed from a sharp to a
supersharp tip and the settings were optimized. New AFM images were recorded with both the TEM grid and
silicon wafer piece of ECI experiment 8. These images [Chap.6.3.2., figure 6.8] were all resized to equal scale, to
gain a more comparable overview of the differences between the two tips and imaging techniques. Figure 7.4
(a) shows the rescaled supersharp tip AFM image on top of a sharp tip AFM image, both on the silicon wafer
piece. Figure 7.4 (b) shows a rescaled 200k magnification TEM image on top of a supersharp tip AFM image. For
comparison of the supersharp tip on the TEM grid and silicon wafer piece or the more detailed original images,
figure 6.8 can be used.
Figure 7.4 (a) reveals that a supersharp tip with associated settings drastically improves the cluster
representation in the image. This creates the possibility to do a valid comparison between the TEM and AFM
Figure 7.3: 200k magnification TEM (top left) and AFM (background) resized to same scale (ECI exp.2)
(a) sharp tip AFM image (top left) and
supersharp tip AFM image (background) both on silicon wafer piece
(b) 200k magnification TEM image (top Left) on supersharp tip AFM image (background)
both on TEM grid
Figure 7.4: TEM and AFM (sharp and supersharp tip) resized to same scale (ECI exp.8)
28 / 38
imaging techniques [figure 7.4 (b)], which failed
between TEM and sharp tip AFM [figure 7.3]. Figure 7.4
(b) shows similar agglomerates and empty spaces
between the particles. However, the particles sizes are
still not the same. Literature research revealed that AFM
imaging is not well suited for accurate examination on
lateral size and shape of single particles, as it will always
yield larger sizes [2], in agreement with figure 7.4 (b).
This is because of the convolution of the AFM tip and
nanocluster radius, as shown in figure 7.5 (a). The tip to
surface interactions on which AFM imaging is based
causes the tip to maintain an equal distance as it scans
the surface. For dense-packed systems without empty
spacing between the particles, this problem does not
arise due to the lack of side contact by the AFM tip, as
shown in figure 7.5 (b).
Even though the convolution problem means that the
width of the displayed particles does not represent the
actual cluster width, the height profile should not be
affected. This is supported by the height scale bars next
to figure 7.4 (a) and (b), which denote a comparable
maximum height of 56.8 nm and 56.5 nm respectively.
The latter means that a height profile analysis is still
possible with the AFM images. By the use of Gwyddion 2.28, a Scanning Probe Microscopy data visualization
and analyses tool, the height distribution was calculated of figure 6.8 (c) [Chap.6.3.2.], as shown in figure 7.6.
This latter AFM image was made with a supersharp tip on the TEM grid of ECI experiment 8 and was considered
the most accurate AFM image made during this research. As can be seen from figure 7.6 the majority of the
deposited copper clusters has a height of approximately 9 nm. The remaining clusters gradually increase in
height, resulting in a plateau, with a small second maximum at 19 nm, which is expected for two stacked
clusters. These maxima are comparable with the approximately 10 nm clusters on the comparable TEM image
on figure 6.8 (a) and the comparison on figure 7.4 (b). Also, this 10 nm size is expected and roughly the same as
the clusters from the HFD experiment series [see Chap.6.2.1., table 7] and therefore are assumed to be valid
height values.
If the charged clusters would have
been deformed due to ECI, this could
have been revealed by a double peak
in the height distribution. A double
peak could have meant that the
(expected substantial amount of)
charged clusters [see Chap.4.1.1.]
were of different height, because of
deformation, disintegration or
embedding in the substrate by the
bias voltage. The other peak would
then have shown the size of the
neutral clusters which are not
affected by an applied bias voltage
and therefore do not deform.
Unfortunately, there are no clear
(a) Problematic convolution between tip and single
particles
(b) Unproblematic convolution between tip and
multiple particles
Figure 7.5: AFM convolution with single and multiple particles. Rt is the AFM tip radius, Rc the cluster radius, r the image radius and h the tip height adjustment [2]
Figure 7.6: height distribution of figure 6.8 (c) [Chap.6.3.2.], supersharp tip AFM on TEM grid (ECI exp.8)
29 / 38
height differences in the height distribution supporting
ECI effects.
Besides convolution another problem occurred during
AFM imaging. Several images showed unnatural
repetition, as shown in figure 7.7, which is impossible
for random cluster deposition. In figure 7.7 this
manifests in small white dots on every clusters. This
familiar problem to AFM imaging occurs when
specimen particles are picked up from the substrate
and stick to the AFM tip. The latter is referred to as
‘double’ tipping, as both the actual tip and picked up
particle act as tips. Initially it was postulated that this
could be an indication of soft landing, which is the
opposite of high speed impact with ECI. But since at
least half the clusters have a neutral charge, this was
rejected.
7.3. Oxygen interference
During the ECI experiment series it became increasingly clear that the system had a leak and undesired oxygen
found its way in. First evidence was found by a colleague working with magnesium as sputtering target
material. TEM imaging of his samples showed that during sputtering magnesium oxide had formed, instead of
the expected magnesium core particles. This indicated that oxygen entered the system during sputtering.
Looking back at table 9 [Chap.7.1.2.], oxygen leakage into the system also looks like an appropriate
explanation, as it would happen over time, regardless of the amount of experiments.
After some more inquiries with colleagues at the University of Groningen, who frequently work with UHV
systems, it was suspected that the blue nylon tubing through which the gasses were supplied is permeable for
oxygen. This permeation can occur because of the oxygen concentration gradient from the outside to the
inside of the nylon tubing. The outer atmosphere contains roughly 21% oxygen and the sputtering gasses
almost none. To exclude the supply of oxygen particles from the argon and helium input, the nylon tubing was
replaced by stainless steel after ECI experiment 6 [see Chap.5.1. and Chap.5.3., table 6]. Also, moisture and
oxygen washers were installed on top of the gas cylinders at the beginning of the gas supply tubing. Because of
these additions sputtering of magnesium became impossible at the common settings. Sputtering experiments
of copper however, still produced particles but with a far lower yield [see Chap.6.3.1., figure 6.6]. These results
lead to the assumption that oxygen leaking into the system aided the sputtering process and particularly
strongly assisted in particle nucleation. It was during this research that the custom build QCM (Quartz Crystal
Microbalance) [25] proved to be a valuable indicator for the workings of the system. It distinctly showed
whether or not sputtering occurred without the need for time consuming TEM imaging.
To check if the Nanosys550 sputtering system was working properly a mechanic from MANTIS Deposition LTD
visited for some testing. He concluded that the system had no defects and suggested to try some depositions
with the Turbo vacuum pumps turned off to increase the operating pressure. Later experiments, through
stainless steel tubing with washers, showed that sputtering with only the aggregation volume Turbo pump
turned on did not give any deposition, but sputtering without any Turbo pump surprisingly did, according to
the QCM. Further research also revealed that any solution against leakage with the use of stainless steel tubing
or gas washers increased the necessary gas input for achieving sputtering. Still for some experiments the
maxima settings of the system did not seem enough. More unexplainable behavior was observed during testing
which seemed related to oxygen. One of those being high sputtering rates for newly placed sputtering
materials accompanied by sparking on the target surface. Again, these results pointed to the fact that oxygen
leakage somehow positively influenced the sputtering process because of its extra partial pressure or chemical
properties.
Figure 7.7: double tipping problem with sharp tip AFM imaging, manifesting in additional small white dots on every cluster (ECI exp.8)
30 / 38
In the end, a solution was found to overcome the leakage problems and sputter without oxygen interference.
By shrinking the orifice aperture, between the aggregation and deposition volume, to a size below the original
5 mm diameter, sputter conditions improved to common system settings. The workings of this change can be
explained as suggested earlier [Chap.4.2.1.]. By shrinking the aperture, the pressure in the aggregation volume
increases, which presumably shortens the mean free path of the gas particles. The latter will increase the
amount of collisions between the particles and promote supersaturation at lower system settings, thereby
encouraging three body collisions for nucleus formation and later on cluster growth.
Further evidence and explanations on the observed behavior because of oxygen leakage came from literature
research. The research done by Marek et al. [30] revealed that even oxygen-to-argon admixture levels of the
order of 1:1000 had a considerable influence on copper sputter. His research was motivated by the repeated
observations of enhanced cluster generation after startup of the source and thus pointed at significant effects
due to contaminations and background gasses from other researchers. Employing a Quadrupole Mass Filter,
Quartz Crystal Microbalance and premixing of oxygen with argon through several mass flow controllers,
enabled precise research on the deliberate admixture of small percentages of oxygen. His work showed that
three distinct modes could be described: clean mode (less than 0.02 sccm O2), low oxygen mode (up to 0.05
sccm O2) and high oxygen mode (more than 0.06 sccm O2), with a flow of 124 sccm argon. During the
experiments the oxygen admixture was raised stepwise to go through all of these modes and back again. It
revealed that increasing oxygen supply increased the amount and size of the deposited copper clusters and
that this transition was reversible. At 0.04 sccm a maximum was reached, presenting a high but instable
Quadrupole collected current and mass-deposition rate on the quartz crystal. Further admixture increase,
decreased the deposition rate. Experimentation was also done with a titanium target. It was found that the
modes for titanium were much higher, with the low oxygen mode between 0.2 and 0.3 sccm. Also, the low
oxygen mode was found to be unstable and shift to high oxygen mode, which did not produce detectable
clusters. In his discussion Marek describes that the oxygen probably interferes on all stages of the sputtering
process: (1) processes on the target, (2) nucleus formation and (3) cluster growth.
With respect to the target, oxidation of the surface will decrease the electron emission and sputter yield. Even
more so, high admixture appeared to poison the titanium target, as sputtering stopped all together. Oxygen
also helps with the seed formation by three body collision as Marek showed that the produced clusters are not
only bigger but also more numerous in low oxygen mode. He additionally notes that metal oxide molecules can
possibly support seed creation and that oxygen-driven microarcing could enhance seed formation. Both the
former and the latter were experienced during this research. The former was seen with the magnesium oxide
particles as the first evidence of oxygen and when the oxygen was removed stopped magnesium sputtering.
The latter was clearly audible and visible as sparking with newly installed target materials. Finally the cluster
growth also experienced changes, as Marek postulates that the binding energy of the clusters may be modified
due to oxygen. If this is the case, as the results support, the balances of all growth processes like coagulation,
coalescence and evaporation are distorted, resulting in the described admixture modes [Chap.3.2].
Although, the HFD results were severely influenced by the oxygen intrusion during the HFD experiment series,
no explanation was found for the absence of results supporting ECI due to oxygen. Experts from MANTIS
Deposition LTD were also contacted for this problem, but to their knowledge, oxygen does not shield the
negative charge of the copper nanoclusters, which would prohibit ECI. This was substantiated by the fact that
they do not encounter oxygen problems with their Nanoshell Coater unit. This additional unit can be placed in-
line after the nanoparticle source on the Nanosys550 Deposition System, between the aggregation volume
(after the exit orifice) and the deposition volume. It generates a second vapor of desired material to coat an
outer shell on the fully grown clusters by electrostatic assembly while passing through this second vapor,
before exiting into the deposition volume. If oxygen would shield the cluster charge, this electrostatic assembly
would not be possible.
31 / 38
8. Conclusions
When comparing the conclusions of Marek’s work with the Half Factorial Design (HFD) experiments done on
the Nanosys550, the majority of the behavior can be explained by oxygen interference [Chap.7.3.]. It seems
plausible that when the oxygen leakage was removed from the system, the sputtering process became harder
to accomplish and the operating windows shifted up, because the helping hand of the oxygen was removed.
The higher settings were not practicable for the system and the more difficult sputtering of magnesium
stopped. The smaller orifice resolved this problem as the increase in pressure lowered these windows to
workable settings [Chap.7.3.]. Furthermore, the insignificant results of the initial HFD are now comprehensible
with the knowledge that oxygen interfered with all steps of the process. With these gained insights a new
Factorial Design would probably result in the expected results with clear significance.
With respect to the Energetic Custer Impact (ECI) experiment series, no evidence was found to support or proof
ECI and therefore cluster charge. Although various literature sources describe the presence of significant
amounts of negatively charged copper clusters, which also seems highly likely because of the energetic nature
of the sputtering process, no corroborating data was obtained. TEM imaging revealed no deformation or
disintegration of cluster upon impact with the substrate and AFM imaging also did not show any remarkable
height differences. The TEM images on ECI [Chap.6.3.1., figure 6.6] do show a yield increase for the high bias
voltages with respect to the low bias voltages, which could suggest charged copper clusters. But even if this
increase is caused by present cluster charge, the kinetic energy gain towards the substrate is not enough to be
called ECI. Even though, AFM imaging on sputtered copper clusters needs more optimization experiments to
generate valid data, it could be useful for height profiling in future research, when convolution is taken into
account. To further examine the charge character of the copper nanoclusters a method to separate the grown
clusters by their charge, as described in literature [6], is advisable.
As a final note, the discovery and recognition of oxygen took quite some time. This is where thorough literature
research, discussion with fellow scientists and a curious mind for noting unusual results could have proven its
value. The problem had been spotted earlier, as TEM images at the beginning of this research showed cubic
particles. Although the group leader professor Kooi, remarked that this could well be a consequence of oxygen
interference during sputtering, it was not recognized as problematic. In retrospect, comparison of these cubic
particles with the work of Olynick et al. [28], on trace oxygen effects on copper nanoparticle size and
morphology, would have shown the same truncated octahedral (cubic in projection) for a controlled O2 leak
rate of approximately 1 Pa.L/s. Further research on the effects of oxygen could have exposed the gravity of the
encountered problems sooner and revealed the appropriate solutions.
32 / 38
9. References
1. Koch, S. (2005). Functionality and dynamics of deposited metal nanoclusters. Groningen: Groningen
University Press. ISBN 90-367-2289-6
2. Hornyak, G.L., Peschel, ST., Sawitowski, Th., Schmid, G. (1998). “Tem, Stm and Afm as tools to study
clusters and colloids,” Micron, vol.29 (no.2/3), pp. 183-190
3. Mazalova, V.L., Soldatov, A.V., (2008). “Geometric and electronic structure of small copper
nanoclusters xanes and dft analysis,” Journal of Structural Chemistry, vol.49, Supplement, pp.S107-
S115
4. Ibrahimkutty, S. (2005). Production, deposition and characterization of metal nanoclusters using a gas
aggregation source. Greifswald: Ernst-Moritz-ArndtUniversität
5. Alivisator, A.P. (2000). Alivisator Group Homepage, www.cchem.berkeley.edu/pagrp/, Retrieved 14
June, 2012, University of California Berkley
6. Haberland, H. Mall, M., Moseler, M., Qiang, Y., Reiners, T., Thurner, Y., (1994). “Filling of micron-sized
contact holes with copper by energetic cluster impact,” Journal of Vacuum Science & Technology A,
vol.12 (no.5), pp. 2925-2930
7. Shin, D.-W., Dong, C., Mattesini, M., Augustsson, A., Mao, S., Chang, C.L., Persson, C., Ahuja, R.,
Nordgren, J., Wang, S.X., Guo, J.-H., (2006). “Size dependance of the electronic structure of copper
nanoclusters in SiC matrix,” Chemical Physics Letters, vol.422, pp. 543-546
8. Haberland, H., Karrais, M., Mall, M., Thurner, Y. (1992). “Thin films from energetic cluster impact: A
feasibility study,” Journal of Vacuum Science & Technology A, vol.10 (no.5), pp. 3266-3271
9. Murai, K., Tokoi, Y., Suematsu, H., Jiang, W., Yatsui, K., Niihara, K. (2008). “Particle Size Controllability
of Ambient Gas Species for Copper Nanoparicles Prepared by Pulsed Wire Discharge,” Japanese
Journal of Applied Physics, vol.47 (no.5), pp. 3726-3730
10. Feiden, P., Stehlé, J., Leygnier, J. (2008). “Growth of gas phase nanoparticles with an accretion
mechanism,” European Physical Journal D, vol.50, pp. 53-60
11. Knauer, W. (1987). “Formation of large metal clusters by surface nucleation,” Journal of Applied
Physics, vol.62 (no.3), pp. 841-851
12. Smirnov, B., Ibrahimkutty, S., Hippler, R. (2007). “Flow of nanosize cluster-containing plasma in a
magnetron discharge,” Physical Review E, vol.75 (no.6), pp. 066402-1(9)
13. Marek, A., Valter, J., Vyskocil, J. (2011). “Cu nanoclusters prepared by magnetron based gas
aggregation technique – effect of magnetic field configuration,“ 30th ICPIG, August 28th – September
2nd, 2011, Belfast, Northern Ireland, UK
14. Flagan, R.C., Lunden, M.M. (1995). “Particle structure control in nanoparticle synthesis from the vapor
phase,” Materials Science and Engineering, vol.A204, pp. 113-124
15. Gracia-Pinilla, M., Martinez, E., Silva Vidaurri, G., Perez-Tijerina, E. (2009). “Deposition of size-selected
Cu nanoparticles by inert gas condensation,” Nanoscale Research Letters, vol.5, pp. 180-188
16. Johnson, R.L. (2002). Masters Series in Physics and Astronomy: Atomic and Molecular Clusters. Taylor
& Francis e-Library (2005), ISBN 0-203-48489-4
17. Briehl, B., Urbassek, H.M. (1998). “Monte carlo simulation of growth and decay processes in a cluster
aggregation source,” Journal of Vacuum Science & Technology A, vol.17 (no.1), pp. 256-265
18. F. J. Giessibl, F.J. (2003). “Advances in atomic force microscopy,” Reviews of Modern Physics, Vol.75,
pp. 949-983
19. Eising, G., Kooi, B.J. (2012). “Cluster evolution and critical cluster sizes for the square and triangular
lattice Ising models using lattice animals and Monte Carlo simulations,” Physical Review B, vol.85
(no.21), pp. 214108-1(12)
20. Granqvist, C.G., Buhrman, R.A. (1976). “Ultrafine metal particles,” Journal of Applied Physics, Vol. 47,
No.5, May 1976
21. Banerjee, A.N., Krishna, R., Das, B. (2008). “Size controlled deposition of Cu and Si nano-clusters by an
ultra-high vacuum sputtering gas aggregation technique,” Applied Physics A, vol.90, pp. 299-303
33 / 38
22. Moseler, M. Rattunde, O., Nordiek, J., Haberland, H. (2000). “On the origin of surface smoothing by
energetic cluster impact: Molecular dynamics simulation and mesoscopic modeling,” Nuclear
Instruments and Methods in Physics Research B, vol.164-165, pp. 522-536
23. Quesnel, E., Pauliac-Vaujour, E., Muffato, V. (2010). “Modeling metallic nanoparticle synthesis in a
magnetron-based nanocluster source by gas condensation of a sputtered vapor,” Journal of Applied
Physics, vol.107, pp. 054309(8)
24. Smirnov, B.M. (2010). Cluster Processes in Gases and Plasmas. Weinheim: Wiley-VCH Verlag GmbH &
Co. KGaA, ISBN 3527409432
25. Momand. J. (2011). “Implementation of a QCM system,” University of Groningen: Nanostructured
Materials and Interfaces group, 22 July, 2011
26. Electron Microscopy ETH Zürich (14 May, 2012). Electron Microscopy, www.microscopy.ethz.ch,
Retrieved 4 July, 2012, ETH Zürich
27. Binnig, G., Quate, C.F. (1986). “Atomic Force Microscope,” Physical Review Letters, vol.56 (no.9), pp.
930-934
28. Olynick, D.L., Gibson, J.M., Averback, R.S. (1996). “Trace Oxygen effects on copper nanoparticle size
and morphology,” Applied Physics Letters, vol.63 (no.3), pp. 343-345
29. Andries, J.P.M., Van Arendonk, R.J.F.M., Bouman, J., Hartsema, J.G., De Vries, A.B. (1992).
Chemometrie II. Groningen: Projectteam Chemometrie, Rijkshogeschool Groningen, Sector Natuur en
Techniek, Afd. HLO/CT
30. Marek, A., Valter, J., Kadlec, S., Vyskocil, J. (2011) “Gas aggregation nanocluster source — Reactive
sputter deposition of copper and titanium nanoclusters,” Surface and Coatings Technology, vol.205,
Supplement 2, pp. S573-S576
34 / 38
Appendix A.1. Experiment design
When operating new equipment of any kind the first task at hand is to familiarize oneself with the basic
operation and characterize the corresponding results. In the case of magnetron sputtering there are several
process variables, which all have specific functional operating windows [Chap.4.2.]. For such a complex process
it is desirable to find the optimum settings for the required results before fulltime usage for long-term scientific
research. The theory describing this practice is called ‘Experimental Optimization’ (EO) [29] and comprises
several tools for obtaining the common settings for operation in an efficient way. With EO the experimental
setup is approached as a ‘black box’ of which the inner mechanisms are unknown, so the settings are directly
related to the results. This is especially useful for magnetron sputtering as it is difficult to evaluate the cascade
of chemical reactions inside the aggregation volume [see Chap.3.] and because of the coinciding effects of
several factors [Chap.4.].
There are two main optimization approaches in EO: sequential and simultaneous. With sequential optimization
the settings of a subsequent experiment are derived by analysis of previous experiments with the aim to find
the appropriate values for the desired response as fast as possible. Examples of Sequential Optimization are:
the Fibonacci method, Simplex- and Steepest Ascent method. In contrast, Simultaneous Optimization focusses
on finding the total behavior of the system and thereby characterizing the influence of every factor, for
example with a: Random-, Factorial- or Central Composite design. With simultaneous optimization a schematic
series of experiments is preformed to obtain the total response function or space. Both methods can be
performed by changing one (univariation) of several (multivariation) factors at same the time. In EO the
process variables and corresponding results are called ‘factors’ and ‘responses’, respectively.
A.1. Factorial Design (FD)
A normal Factorial Design (FD) is probably the most common and intuitive approach for scientific research,
although not always named this way. Essentially, the range of every factor is divided in equal steps and each
combination of these steps and of all factors is used during experimentation. For a FD this is described as the
following. First the ranges of all adjustable process factors are plotted in an N-dimensional factor space, with
one dimension for every factor for N factors in total. Then the factor space is divided in equally sized (hyper-)
cubes. The amount of desired data point per factor dictates the amount of cubes. If for an arbitrary factor only
two data points are desired, the range of that factor will be covered by two cubes. The center coordinate
values of each of these (hyper-) cubes are then used as the factor settings for a single experiment. The results
of all of these center value experiments together comprise the response space.
Even though a higher amount of data point per factor will yield a more accurate response space in the end, this
approach has its shortcoming as the amount of required experiment grows exponentially with the amount of
data points per factor. From this impracticality, the 2k FD with only a high and low data point (i.e. setting) per
factor is commonly used for initial characterization. It requires only 2k experiments for k factors, if no
duplications are incorporated. To simplify the notation during experimentation the high settings are
conventionally denoted +1 and the low setting -1. If three settings are desired a middle setting, denoted 0, can
also be added. Although more intuitive for quantitative variables this system works equally well for qualitative
variables, like different gasses A (as +1) and B (as -1).
When all adjustable factors are known, a schematic experiment series can be constructed. An exemplary
experiment series with only three process factors and their experiment responses is shown in table A.1. For
clarity to the reader this table has a systematic order, as every factor is set to high and low in a systematic
fashion. During experimentation it is advised to maintain a random sequence of these experiments. This can
potentially reduce or reveal any influences of non-controllable factors like buildup of charge or material, in the
final response. In the exemplary table A.1 the adjusted factors are denoted by unicase letters (a, b, c) and their
resulting responses as capitals. With this approach not only main effects (A, B, C) from only one changed factor
can be investigated, but also their duo interaction effects (AB, AC, BC) and triple interaction effect (ABC).
35 / 38
Table A.1: Example of experiment series by Factorial Design Experiment # Combination Effect A Effect B Effect C Response
1 (0) - - - y1 = 62 2 a + - - y2 = 68 3 b - + - y3 = 62 4 ab + + - y4 = 66 5 c - - + y5 = 61 6 ac + - + y6 = 64 7 bc - + + y7 = 65 8 abc + + + y8 = 69
As for the quantification and qualification of the experimental data in general, one has to be careful in drawing
conclusions. Calculation on these responses of a FD will only yield approximations of the observed effects,
because of the absence of duplicates and limited data points per factor.
For the quantitative calculation of the main, duo or triple effects the experiments are grouped in pairs with a
high and low setting of the effect under consideration while the other settings are kept equal. Then, the paired
experiments are subtracted and the results of all groups are averaged. Table A.2 shows the examples for a
main (A) and duo interaction effect (AB). For the effect AB, notice that the high and low B settings are
separated in a high and low group as described.
Table A.2: Example of calculating contribution of separate effects
Main effect A Duo interaction effect AB
Setting A Setting A Setting B Setting C + - Difference Setting B Setting C + - Difference
- - 68 62 6 (=y2-y1) + - 66 62 4 (=y4-y3) + - 66 62 4 (=y4-y3) + + 69 65 4 (=y8-y7)
- + 64 69 3 (=y6-y5) Average 4
+ + 69 65 4 (=y8-y7) - - 68 62 6 (=y2-y1) - + 64 61 3 (=y6-y5)
Average 4.25 Average 4.5
From table A.2 the increase of response due to main effect A is the same as the average of 4.25. For the duo
interaction effect AB, the change in response due to altering settings is by definition half the difference
between these two averages, AB = ½ . (4 - 4.5) = -0.25. If there would not be an interaction effect AB between
A and B, the averages would be the same and result of effect AB would be zero. Following the same system the
triple interaction effect ABC can also be calculated. For this the effect AB is calculated separately for the high
and low setting of C. So ½ [(y4 - y3) - (y2 - y1)] = ½ (4 - 6) = -1 for low C and ½ [(y8 - y7) - (y6 - y5)] = ½ (4 - 3) = -
0.5 for high C. Then, the same definition of half the difference applies, so effect ABC = ½ (0.5 - -1) = 0.75.
Table A.3: Sign table for simplifying calculations Experiment # A B C AB AC BC ABC Response
1 - - - + + + - y1 2 + - - - - + + y2 3 - + - - + - + y3 4 + + - + - - - y4 5 - - + + - - + y5 6 + - + - + - - y6 7 - + + - - + - y7 8 + + + + + + + y8
To simplify the calculation a ‘sign table’ can be drawn (table A.3). In such a table the product of the + and -
signs of the factor settings result in the sign of the interaction effects (e.g. Experiment 2: A ∙ B = AB, +1 ∙ -1 = -1).
Now calculation of any effect can easily be done. First group the experiment with high settings (+) and those
36 / 38
with low settings (-) according to the table and second subtract the average of the negative group from the
average of the positive group, to get the response of that particular effect. This can be denoted as:
. For example:
It has to be emphasized that the interaction effects are the degrees to which the underlying main effects are
not simply additives. So even though main effect A and B lead to their own change in response, together they
cannot simple be added, but their interaction also has to be considered. This is also true for the triple
interaction ABC which estimated the degree to which the interaction effect AB and main effect C are not simply
additive.
Even though the above illustrated calculations are simple, they can be quite time consuming and with more
thorough experiments the amount of data points will grow exponentially. To overcome this problem the Yates
algorithm can be used, which will not be discussed here.
A.2. Half Factorial Design (HFD)
For the characterization of copper nanoparticle DC sputtering on the Mantis Nanosys550 Deposition System an
Half Factorial Design (HFD) was chosen because of its small amount of initially required experiments and
therefore crude but fast characterization of the system. In a normal 2k FD there is a doubling of required
experiment with every additional factor (k+1), also the higher order interaction effects will increase (e.g. k = 4
gives quadruple interaction ABCD). These higher order effects, usually from triple interactions up, are rarely
significant and so produce superfluous information. But this superfluous information can be used differently,
by confounding (i.e. blending) the higher orders effects with the main effects in a linear combination. This
allows for an incomplete FD, which reduces the required experiment to a Half or Quarter 2k Factorial Design
(HFD or QFD), with 2k-1
or 2k-2
experiments, respectively. With a HFD a four factor system can be characterized
with only eight experiments. The drawback of this drastic reduction in experiments is that with these only the
main effects can be positively quantified. For the interaction effects a HFD cannot give a quantitative response,
but only an indication of existence of an interaction if the response is relatively high. Also, if the HFD results in
no significant effects, this does not mean that there are not any and additional experiments become
mandatory. This proves that the used sample size of experiments has been too small. As a consequence of the
confounding, the results from the experiment are no longer pure responses and therefore are referred to as
‘aliases’.
When constructing a HFD, the fourth factor, denoted D, will be confounded with the highest interaction effect
ABC, so D = ABC. To do so, the factor D is set to the same high/low settings as the column of ABC in table A.3
and is set alongside A, B and C before starting experimentation. Now to find the individual responses the
confounding has to be analyzed. This can be done by incorporating the extra factor D in a sign table like table
A.3 and completing it with the extra (interaction) effects: D, AD, BD, CD, ABD, ACD, BCD and ABCD. As it turns
out, some of the columns of this new table will have exactly the same sequence, implying that they’re
associated effects are confounded. These equals can be combined in three groups: (1) main effects with triple
interactions A + BCD, B + ACD, C + ABD, D + ABC; (2) double duo interactions AB +CD, AC + BD, BC + AD; and (3)
the quadruple interaction: ABCD + the average, as shown in table A.4 (column Alias). Because we stated that
triple interactions and higher orders are mostly insignificant with respect to the main effects, they can be
omitted, thereby revealing the main effects and average (experiments 2, 3, 5 and 8). With respect to the duo
interaction effects, a large alias will not reveal any absolute effects, but is a good indication for underlying
phenomena.
As mentioned above, the drawback of this HFD is that some additional experiments have to be done for the
duo and triple interaction effects if no clarifying foreknowledge is present. Seemingly, this would reduce the
biggest advantage of a HFD. Still there can be an advantage over a normal FD with 24 experiments depending
on the initial outcome and desired level of characterization. The main effects can be found faster and the
amount of additional experiments can be chosen according to the desired interactions.
37 / 38
Revealing the duo and triple interaction effects can be accomplished by composing a new sign table with the
difference that D = - ABC. This will generate a complementary design with minus signs between the
confounded aliases, as shown in table A.4 (column Complementary Alias). Addition and subtracting of the
complementary aliases with the initial aliases will reveal the underlying responses (e.g. Experiment 4: ½ [(AC
+BD) + (AC - BD)] = AC and ½ [(AC +BD) - (AC - BD)] = BD). So with the use of these complementary experiments,
one can also find the desired duo, triple and quadruple interaction effects.
Table A.4: Half Factorial Design with complementary experiments Experiment # A B C D D* Alias Complementary
Alias* Reveals
1 - - - - + ABCD + average ABCD - average Average (, ABCD*) 2 + - - + - A + BCD A - BCD A (, BCD*) 3 - + - + - B + ACD B - ACD B (, ACD*) 4 + + - - + AB + CD AB - CD AB (, CD*) 5 - - + + - C + ABD C - ABD C (, ABD*) 6 + - + - + AC + BD AC - BD AC (, BD*) 7 - + + - + BC + AD BC - AD BC (, AD*) 8 + + + + - D + ABC D - ABC D (, ABC*)
* Additional experiment required for complementary set
A.3. Center Point Experiments (CPEs)
The main reason of characterizing a system as efficient as possible by as few experiments as possible does have
its price with respect to its quality. A shortcoming of which every type of FD with only one high and low setting
suffers is the inability to yield non-linear relations. Also, the desire to minimize the amount of experiments can
result in the absence of duplications and thereby burdens the reliability of the single experiments. To
compensate for these inherent quality loses Center Point Experiments (CPEs) can be added to the experiment
series. For these CPEs the factors of the FD are all set to their center value between the high and low setting
and denoted with a row of zeros in the sign table. Preferably, the experiment series contains several of these
CPEs at strategic positions to increase their utility.
Adding CP’s has several advantages to increase the quality of the overall research. The first is that from the
CPEs a baseline can be constructed to which the other experiments can be compared. As for a stable system
the response of every CPE should be the same, which results in a flat baseline when plotted. If the baseline has
a slope it can be concluded that the system exhibits accumulating or degrading phenomena, which introduce a
time dependent error in the experiments. The same holds for a curvature in the baseline. This can also reveal
non-controlled (noise) factors, which have to be eliminated or compensated for. In addition, the CPEs can
reveal non-linear behavior of a factor if their responses do not lie in the middle between the high and low
responses of the examined factor. Furthermore, introducing several CPEs generates comparable data points
from which a variation or standard deviation can be calculated. The latter being the characteristic error of the
system and can thus be used to evaluate the significance of the responses.
Again, one has to be cautious with interpreting the responses of the CPEs. For example, if the baseline is not a
horizontal line it can be used to detect non-controlled interfering effects, but probably not for estimating non-
linear effects.
38 / 38
Acknowledgments
First and foremost I would like to thank my professor Bart Kooi and supervisor Gert ten Brink.
I have a high appreciation for people who work hard, but this is especially true for those who make it seem
easy and still take the time to listen to your questions and give complete, helpful answers. Bart is one of those
people. Not only did he help with his detailed knowledge, but he also contributed with his unmatched TEM
skills. My thankfulness grew even more after my master’s research when he found a company for my industrial
internship. The reader should know that it took me some extra years to get through my study of Applied
Physics and Bart gave me the opportunity to continue without any further delay nor financial setbacks. Even
when my first internship did not work out as we hoped he found a second one within a week, which I liked so
much that I applied for a job and signed my first contract even before my graduation. For me this will always be
one of those valuable moments in life, where my future took shape just the way I hoped.
During my research I was assigned to Gert as my supervisor, with which I was very pleased. I know Gert from a
previous project and he is always prepared to help. We worked closely together during my research and he
really showed that he wanted me and the group to grow to a higher level. He sometimes joked: “I do not have
the biggest brain,” but his organized way of experimenting and documenting proved invaluable. He taught me
that a scientist should not always change his or her experiments when he or she encounters strange behavior
and that every experiment produces some form of useful data. The latter really showed when we encountered
the oxygen intrusion problem as described in this thesis. Still, with the generous addition of some of Gert’s data
I could write a complete thesis. The resulting chapters on HFD show that, despite his own joke, Gert really is a
true scientist.
I also like to thank assistant professor George Palasantzas and PhD Gopi Krishnan for their help.
I would like to thank George for his confidence in me when operating the sputtering system and AFM. Also,
when the sputtering system was subject to oxygen problems his knowledge and experience proved invaluable.
He furthermore aided in the AFM analysis, which turned out not to be as trivial as expected. At the beginning
of my research I needed to get used to his Greek enthusiasm, which can really blow you away, but it does not
take very long to discover his kind character and appreciate his big bearded smile, which is very catchy.
Simultaneous to my research, Gopi also worked with the sputtering system. It was very useful to have
somebody who experienced the same problems. This gave the opportunity to learn from each other and share
ideas on the systems behavior, which resulted in a very pleasant cooperation.
In addition I want to thank Orcun Ergincan, Arno Kroezen en Robin de Wit for their help with the AFM. They
patiently showed Gert and me how to operate it on our own and did not mind repeating it several times for a
better learning experience. As can be seen in this thesis, this gave some valuable new insights in how to scan
the substrates and analyze the results, which proved to be a real asset in the overall research.
Finally, I want to thank the rest of the NMI group for a very pleasant time. I really appreciate the tea-time-
discussions we had on everyone’s research and any other topic. Everybody’s genuine interest and attention for
one another really showed the foundation of modern research. We had some healthy differences of opinion
and also a lot of laughs, which were all bound by a coinciding interest for science and mutual respect.
Dame en heren, bedankt!
Arjan ‘Bijl’ Bijlsma